Paper deep dive

Breaking Homogeneity: Diversifying Persona Sets for Creative LLM Outputs

Sang Bin Moon, Nicole Cho, Daniel Borrajo, Sumitra Ganesh, Abolfazl Hashemi

Year: 2026Venue: arXiv preprintArea: cs.CLType: PreprintEmbeddings: 278

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 96%

Last extracted: 9/28/2026, 3:07:46 AM

Summary

This paper addresses the homogeneity of Large Language Model (LLM) outputs by proposing persona diversification as a set-level conditioning mechanism. The authors define a design space crossing selection vs. generation and space-filling vs. frontier-seeking diversity, instantiating four methods: Coverage, Dispersion, Uniform-Coverage MCMC (UC-MCMC), and evolutionary TextGrad. Empirical evaluations on the Alternative Uses Task (AUT), Infinity-Chat, and Divergent Association Task (DAT) demonstrate that evolutionary persona generation significantly increases response diversity, originality, flexibility, and holistic creativity compared to task-only prompting and random personas, while maintaining high validity.

Entities (17)

Abolfazl Hashemi · person · 99%Daniel Borrajo · person · 99%JP Morgan AI Research · organization · 99%Nicole Cho · person · 99%Purdue University · organization · 99%Sang Bin Moon · person · 99%Sumitra Ganesh · person · 99%Alternative Uses Task · benchmark · 98%Divergent Association Task · benchmark · 98%Infinity-Chat · benchmark · 98%Coverage · algorithm · 95%Dispersion · algorithm · 95%Evolutionary TextGrad · algorithm · 95%Uniform-Coverage MCMC · algorithm · 95%Groupthink · concept · 92%RLHF · method · 90%TextGrad · method · 90%

Relation Signals (11)

Sumitra Ganesh → affiliatedwith → JP Morgan AI Research

confidence 99% · Sumitra Ganesh ... Affiliation: J.P. Morgan AI Research

Sang Bin Moon → affiliatedwith → Purdue University

confidence 99% · Sang Bin Moon Affiliation: School of Electrical and Computer Engineering Affiliation: Purdue University

Nicole Cho → affiliatedwith → JP Morgan AI Research

confidence 99% · Nicole Cho ... Affiliation: J.P. Morgan AI Research

Daniel Borrajo → affiliatedwith → JP Morgan AI Research

confidence 99% · Daniel Borrajo ... Affiliation: J.P. Morgan AI Research

Abolfazl Hashemi → affiliatedwith → Purdue University

confidence 99% · Abolfazl Hashemi Affiliation: School of Electrical and Computer Engineering Affiliation: Purdue University

Dispersion → implementsprinciple → Frontier-seeking diversity

confidence 95% · the frontier-seeking principle is instantiated by Dispersion and evolutionary TextGrad

Coverage → implementsprinciple → Space-filling diversity

confidence 95% · The space-filling principle is implemented by Coverage and UC-MCMC

Evolutionary TextGrad → usedin → Alternative Uses Task

confidence 95% · On AUT, evolutionary persona generation increases response diversity by 78.8%

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Language models often produce homogeneous responses to open-ended tasks; such homogeneity can spawn groupthink-the convergence of ideas toward a singular and potentially suboptimal decision. We formulate persona diversification as a set-level conditioning problem and study two orthogonal design choices: selecting versus generating personas, and space-filling versus frontier-seeking diversity. We instantiate this design space with four methods spanning coverage and dispersion subset selections, uniform-coverage sampling, and evolutionary persona generation. Evaluations on the Alternative Uses Task (AUT), Infinity-Chat, and Divergent Association Task (DAT) show the benefits of the proposed methods across tasks and creativity objectives. On AUT, evolutionary persona generation increases response diversity by 78.8%, originality by 26.1%, flexibility by 49.5%, and holistic creativity by 13.9% over task-only prompting, while maintaining 98.5% validity; on Infinity-Chat, it nearly doubles persona-induced response separation relative to random personas. Moreover, evolutionary personas compose with creativity-optimized prompting, further increasing its response diversity by 18.6% and creativity by 6.3%. These results establish persona-set geometry as a task-agnostic mechanism for eliciting divergent LLM outputs, and support persona diversification as a reusable complement to prompt optimization.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

274,445 characters extracted from source content.

Expand or collapse full text

Breaking Homogeneity: Diversifying Persona Sets for Creative LLM Outputs Sang Bin Moon Affiliation: School of Electrical and Computer Engineering Affiliation: Purdue University Affiliation: West Lafayette, IN 47909, USA Email: moon182@purdue.edu Nicole Cho Daniel Borrajo & Sumitra Ganesh Affiliation: J.P. Morgan AI Research Affiliation: New York, NY 10172, USA Affiliation: nicole.cho, daniel.borrajo, Email: sumitra.ganesh@jpmorgan.com Abolfazl Hashemi Affiliation: School of Electrical and Computer Engineering Affiliation: Purdue University Affiliation: West Lafayette, IN 47909, USA Email: abolfazl@purdue.edu Abstract Language models often produce homogeneous responses to open-ended tasks; such homogeneity can spawn groupthink—the convergence of ideas toward a singular and potentially suboptimal decision. We formulate persona diversification as a set-level conditioning problem and study two orthogonal design choices: selecting versus generating personas, and space-filling versus frontier-seeking diversity. We instantiate this design space with four methods spanning coverage and dispersion subset selections, uniform-coverage sampling, and evolutionary persona generation. Evaluations on the Alternative Uses Task (AUT), Infinity-Chat, and Divergent Association Task (DAT) show the benefits of the proposed methods across tasks and creativity objectives. On AUT, evolutionary persona generation increases response diversity by 78.8%, originality by 26.1%, flexibility by 49.5%, and holistic creativity by 13.9% over task-only prompting, while maintaining 98.5% validity; on Infinity-Chat, it nearly doubles persona-induced response separation relative to random personas. Moreover, evolutionary personas compose with creativity-optimized prompting, further increasing its response diversity by 18.6% and creativity by 6.3%. These results establish persona-set geometry as a task-agnostic mechanism for eliciting divergent LLM outputs, and support persona diversification as a reusable complement to prompt optimization. 1 Introduction Homogeneity in Large Language Model (LLM) outputs, specifically those that pertain to a narrow subset of WEIRD (Western, Educated, Industrialized, Rich, and Democratic) responses, severely limits the potential of LLMs (Anthis et al., 2025) and engenders a problematic artificial hivemind (Jiang et al., 2026). This homogeneity is known to be spawned by the RLHF fine-tuning pipeline endemic to frontier models (Kirk et al., 2024; West & Potts, 2025); such homogeneity not only stifles creativity but also creates systematic risks of groupthink. For example, relying on LLMs for investment recommendations can engender similar or identical advice across independent teams - which results in compounded risk (IMF, 2024; BIS, 2024). Conditioning a language model on a persona, defined as the societal, biological, and personal traits of the human user, has been presented as a potential lever for steering behavior and eliciting distinct perspectives (Choi & Li, 2024; Ge et al., 2024). However, empirically measuring the impact of diverse personas on the model’s creative ability to generate divergent solutions for the same task is relatively understudied (Paglieri et al., 2026). We observe three gaps in current research: first, the limited breadth of evolutionary methods that truly make a set of personas diverse; second, the underexplored notions of what actually makes a set of personas diverse and a set of outputs creative; and third, the understudied realm of whether a set of diverse personas can trigger the model’s divergent trajectories to generate more creative outputs. In this work, we utilize persona as an explicit, optimizable variable and hypothesize that a set of truly diverse personas can successfully trigger the alternative, divergent trajectories embedded within the model that would otherwise remain inactivated. Moreover, we hypothesize that divergent gains compound when a set of personas is explicitly diversified, ensuring that the induced responses spread across many regions at once, for the same task. Thus, our focus is threefold: what makes a set of personas diverse, how can we induce that diversity with guarantees on the resulting persona-side spread, and does the spread ultimately improve creativity? Figure 1: Persona diversification design space. Two principles of diversity (columns) are each instantiated at two levels of control (rows). The space-filling principle is implemented by Coverage and UC-MCMC, whereas the frontier-seeking principle is instantiated by Dispersion and evolutionary TextGrad. Three algorithms have formal guarantees for persona-space objectives and sampling targets, while evolutionary generation is empirical. To answer these questions, we organize persona diversification along two axes (Figure 1): selection from a fixed candidate pool versus generation beyond it, and space-filling coverage versus frontier-seeking separation. Crossing these two axes yields four methods: Coverage and Dispersion for selection, and Uniform-Coverage MCMC (UC-MCMC) and evolutionary TextGrad for generation (Yuksekgonul et al., 2025). Coverage and Dispersion achieve global optima for their population objectives, UC-MCMC converges asymptotically to an equal-cell target over reachable cells, and evolutionary generation is evaluated empirically. Our central finding is that the performance gains are driven by the specific structural geometry of the persona sets, not merely the presence of a conditioning variable. Across the benchmarks, greater persona-set dispersion is positively associated with greater response dispersion. By quantifying this displacement against the model’s unconditioned output, we demonstrate the specific advantage a persona-conditioned query provides over a standard task-only prompt: namely, increased creativity and a meaningful departure from the model’s narrow default behavior. Our diversification objectives depend on persona geometry rather than downstream task performance, allowing the same persona sets to be reused across tasks. Persona conditioning supplements prompt engineering and can be combined with advanced prompting and reasoning scaffolds. Contributions. ▶ A design space for inducing persona diversity that crosses two levels of control (selection vs. generation) with two diversity principles (space-filling vs. frontier), unifying four methods under one framework (Figure 1). The construction is task-agnostic and compatible with prompt optimization and reasoning scaffolds. ▶ On the selection axis, exact algorithms for both principles, maximal-covering location Coverage and max–min Dispersion, with global-optimality guarantees. ▶ On the generation axis, two complementary methods: a Metropolis-within-Gibbs sampler with provable asymptotic convergence to an equal-cell mixture over reachable semantic cells, and an evolutionary TextGrad procedure that searches for isolated, low-density personas. ▶ Empirical evidence that persona diversification can improve response creativity across three benchmarks, including comparisons with task-only, random-persona and prompt engineering baselines. Figure 2: Persona conditioning diversifies Alternative Uses Task responses. Left: Example uses for a tin can under task-only prompting (top) and persona conditioning (bottom), illustrating a broader range of ideas. Right: Mahalanobis-whitened PCA projection of response embeddings across persona conditions, with convex hulls showing their semantic spread. 2 Problem Formulation While a single persona provides pointwise conditioning, it cannot dictate the structure of a broader collection of viewpoints. To address this set-level challenge, we define the persona diversification problem. The objective is to design a collection of persona-conditioned queries that hold the task prompt and base response model constant, isolating the persona set as the sole optimization variable. We formulate this shared problem and its underlying semantic geometry below. 2.1 The Set-Level Conditioning Problem Let T be the text-string space and Ω⊆ be the specialized persona-string space. A persona is defined as a short, natural-language description p∈Ωp∈ , and a persona condition is a set S of fixed cardinality ||=k|S|=k. For a task prompt x∼x and a frozen downstream response model GθRG_ _R, the model generates a response conditioned on p as yp∼GθR(⋅∣x,p),p∈,y_p G_ _R(\,· x,p), p , (1) where p=∅p= denotes the persona-free reference baseline. Holding x and GθRG_ _R fixed, the downstream goal is for the induced response collection yp:p∈\y_p:p \ to span a broader, more creative range of valid, useful outputs than task-only prompting or an unoptimized persona set. Crucially, our optimization variable and evaluation target are deliberately distinct. Our algorithms construct persona sets based entirely on frozen persona-space geometry, whereas creativity is evaluated strictly on the induced response-space outputs. We do not assume that persona-space distance monotonically transfers to response-space distance, novelty, or quality. Consequently, the guarantees we provide apply solely to the persona-space; whether this persona diversification reliably translates to improved response creativity remains an empirical hypothesis evaluated in Section 4. 2.2 Semantic Geometry Let =pii=1N⊆ΩP=\p_i\_i=1^N_P denote a generic finite candidate population, with [N]=1,…,N[N_P]=\1,…,N_P\. Selection operates on the baseline population 0P_0, while generation expands it into TP_T before downstream selection. Let ϕ:→ℝdφ:T ^d denote the full, unit-normalized output of a frozen text encoder, which supports Matryoshka Representation Learning (MRL) (Kusupati et al., 2022). For Mahalanobis distance, we use the prefix ϕ~(p)=ϕ(p)1:dM φ(p)=φ(p)_1:d_M without renormalizing after truncation. This encoder induces persona-space dissimilarities dmPd_m^P, indexed by m∈cos,2,Mahm∈\cos,2,Mah\, as dcosP(p,p′) d_cos^P(p,p ) =1−ϕ(p)⊤ϕ(p′)‖ϕ⁡(p)‖2‖ϕ⁡(p′)‖2, =1- φ(p) φ(p )\|φ(p)\|_2\|φ(p )\|_2, (2) d2P(p,p′) d_2^P(p,p ) =‖ϕ⁡(p)−ϕ⁡(p′)‖2, =\|φ(p)-φ(p )\|_2, (3) dMahP(p,p′) d_Mah^P(p,p ) =(ϕ~(p)−ϕ~(p′))⊤Σ^P−1(ϕ~(p)−ϕ~(p′)). = ( φ(p)- φ(p )) _P^-1( φ(p)- φ(p )). (4) Here, Σ^P _P is the sample covariance of the truncated embeddings ϕ~(p):p∈0\ φ(p):p _0\, regularized by adding εIdM I_d_M, with ε=10−6 =10^-6 and dM=128d_M=128. Both finite-pool selectors rely on the pairwise dissimilarity matrix and its realized spectrum Dm,ijP D^P_m,ij =dmP(pi,pj), =d_m^P(p_i,p_j), mP _m^P =(Dm,ijP)i,j=1N, = (D^P_m,ij )_i,j=1^N_P, (5) Λm() _m(P) =Dm,ijP:i,j∈[N], =\D^P_m,ij:i,j∈[N_P]\, Lm L_m =|Λm()|, =| _m(P)|, (6) Let r denote a unique dissimilarity value (or radius) drawn from this spectrum. The LmL_m distinct values of Λm() _m(P), including 00, are sorted in ascending order such that rm,(1)<⋯<rm,(Lm)r_m,(1)<·s<r_m,(L_m). 3 Diversification Algorithms Building on the semantic geometry established in Section 2, this section structures our algorithms as a progression in both control and exploratory reach. We begin with Selection operators (Section 3.2) that enforce space-filling and frontier diversity within a fixed candidate population. To break the performance ceiling imposed by this finite pool, we escalate to open-ended Generation methods (Sections 3.3 and 3.4), actively expanding the persona support through a space-filling MCMC sampler and a frontier-seeking evolutionary TextGrad search. To maintain narrative flow, all formal proofs and extended mathematical formulations are deferred to Appendix B and C. 3.1 Diversification Design Space Our framework is organized along two independent axes, as illustrated in Figure 1. Selection extracts a subset ⊆S of size k from a frozen candidate pool P, whereas generation expands the baseline pool 0P_0 with new valid personas before applying a downstream selection protocol to obtain the final k-persona condition. We analyze generation and downstream selection separately because guarantees for the generation process do not automatically transfer to the selected subset. 3.2 Diverse Persona Selection: Coverage and Dispersion To extract a k-sized subset S from a finite candidate pool P, we operationalize the space-filling and frontier principles via two exact selection algorithms. For the space-filling principle, we formulate selection as thresholded coverage. A persona pip_i is considered covered if its distance to the nearest selected exemplar in S is at most a radius r. The fraction of the population covered evaluates to Fcov,m(;,r)=1N∑i=1Nminpj∈dmP(pi,pj)≤r.F_cov,m(S;P,r)= 1N_P _i=1^N_P 1 \ _p_j d^P_m(p_i,p_j)≤ r \. (7) To avoid metric-specific scaling from manually tuning r, we adaptively calibrate the smallest radius rm⋆r _m at which some subset successfully covers the target of ncov=⌈ccovN⌉n_cov= c_covN_P personas, written as rm⋆(,ccov)=min⁡r≥0:max⊆||=k⁡NFcov,m(,,r)≥ncov.r _m(P;c_cov)= \r≥ 0: _ subarraycS \\ |S|=k subarrayN_P\,F_cov,m(S;P,r)≥ n_cov \. The algorithm lexicographically attains this minimal radius, then breaks ties among radius-optimal subsets by maximizing the covered count at the threshold. Since Fcov,mF_cov,m changes only at the discrete values of the pairwise spectrum Λm() _m(P), finding r⋆r reduces to an exact discrete search over the sorted distances. For a target coverage count ncovn_cov, the greedy solution Gm(r)≥ncovG_m(r)≥ n_cov certifies feasibility of a trial radius r and establishes the upper bound r⋆≤r ≤ r. By searching over the sorted distances, the upper bound becomes tighter. At the same time, the classical submodular guarantee bounds the optimal covered count from above as nopt≤⌊Gm(r)/(1−1/e)⌋n_opt≤ G_m(r)/(1-1/e) (Nemhauser et al., 1978). If ncov>⌊Gm(r)/(1−1/e)⌋n_cov> G_m(r)/(1-1/e) , the radius is certified infeasible, establishing the lower bound r⋆>r >r. These certificates narrow the bracket on r⋆r , efficiently reducing the number of exact-solver calls. This iterative algorithm (Algorithm 1) solves both the smallest feasible radius r⋆r and the maximum coverage at r⋆r . Theorem 1 (Exact Calibrated Coverage). In exact mode, provided every radius feasibility probe is certified, the adaptive search algorithm (Algorithm 1) terminates and returns a globally minimal radius r⋆=rm⋆(,ccov)r =r _m(P;c_cov), together with a subset acov,m⋆(,ccov)S _acov,m(P;c_cov) of cardinality k satisfying the maximum coverage at r⋆r . For every ⊆S with ||=k|S|=k and 1≤k≤N1≤ k≤ N_P, Fcov,m(acov,m⋆(,ccov),,r⋆)≥Fcov,m(,,r⋆).F_cov,m\! (S _acov,m(P;c_cov);P,r )≥ F_cov,m(S;P,r ). In particular, its covered count at r⋆r is at least ncovn_cov. The full Integer Linear Program formulation, algorithmic pseudocode, and supporting lemmas and proofs are deferred to Appendix B.1. On the other hand, the frontier principle prioritizes structural separation. The discrete max–min p-dispersion problem (Erkut, 1990) maximizes the minimum dispersion (the smallest pairwise dissimilarity within a subset) over all size-k subsets of the candidate pool. For ||≥2|S|≥ 2, the minimum dispersion objective is defined as Fdisp,m():=minp,p′∈p≠p′⁡dmP(p,p′).F_disp,m(S):= _ subarraycp,p \\ p≠ p subarrayd_m^P(p,p ). (8) Since finding a max–min diverse subset is generally NP-hard, our exact method again exploits the discrete topology of the solution space. Note that the optimal dispersion value must equal exactly one of the pairwise distances in the sorted spectrum Λm() _m(P), so we leverage the threshold-graph equivalence between max–min diversity and clique feasibility (Della Croce et al., 2009). The algorithm builds a graph by inserting edges in descending order of pairwise dissimilarity, transforming the optimization into a series of Boolean structural checks. After a new edge pa,pb\p_a,p_b\ is added, the algorithm searches for a (k−2)(k-2)-clique in their mutual neighborhood. If found, combining it with the endpoints yields a complete size-k clique; if not, the algorithm checks for the next pair. Since the algorithm searches the descending discrete distance spectrum sequentially, the subset returned upon the first clique completion is guaranteed to be optimal. Theorem 2 (Exact Max–Min Dispersion). For a finite candidate pool P, 2≤k≤||2≤ k≤|P|, and fixed symmetric nonnegative dissimilarities dmPd_m^P, the incremental-clique algorithm (Algorithm 2) terminates and returns a size-k subset disp,m⋆()S _disp,m(P) satisfying Fdisp,m(disp,m⋆())=max⊆||=k⁡Fdisp,m().F_disp,m(S _disp,m(P))= _ subarraycS \\ |S|=k subarrayF_disp,m(S). 3.3 Uniform-Coverage MCMC Persona Generation To break the diversity ceiling imposed by a finite baseline pool 0P_0, we transition to generation beyond 0P_0. Uniform-Coverage MCMC (UC-MCMC) applies the space-filling principle directly to the generation process by targeting equal probability mass across reachable cells of a fixed partition of equal-area directional cells. Within each cell, a frozen Large Language Model acts as the base law, ensuring the sampled personas remain linguistically plausible and structurally valid while the MCMC framework enforces the geometric uniformity. Since the total number of reachable cells is unknown in advance, UC-MCMC employs a hindsight-spawning random scan that maintains a dynamically growing population of parallel Markov chains, with one chain for each discovered cell. Each iteration selects an active cell uniformly at random and proposes a new persona using the frozen LLM. If the generated persona lands in the currently selected cell, it undergoes a standard Metropolis–Hastings (MH) correction to preserve the within-cell target law. If the proposed persona lands in a previously empty, undiscovered cell, it serves as a hindsight seed that activates the new cell and initializes a new parallel chain. Theoretical Guarantee. By consolidating the cell-kernel reversibility and eventual discovery properties, we establish three formal guarantees for the emitted sequence. Theorem 3 (Anytime Cell Uniformity and Target Convergence). Fix a deterministic validity rule valval, a base generation law b, and a cell map c:Ω→[M]c: →[M] assigning each persona to one of M cells. The encoder, whitening transform, and partition defining c are frozen. Assuming b⁡(p)>0b(p)>0 for every valid persona, we define the valid state space and base mass of cell j as Ωj=p∈Ω:val(p)=1,c(p)=j,Bj=b(Ωj). _j=\p∈ :val(p)=1,\ c(p)=j\, B_j=b( _j). Let ℛ=j:Bj>0R=\j:B_j>0\ be the set of reachable cells and πj=b(⋅∣Ωj) _j=b(\,· _j) their within-cell target laws. Suppose Algorithm 3 is run from a fixed, nonempty active set ℐ0⊆ℛI_0 . With all proposal kernels and mixture weights frozen, the global proposal draws independently from b with a fixed probability ω0>0 _0>0. Let ℐt−1⊆ℛI_t-1 denote the active set after iteration t, PtemitP_t^emit the emitted persona, and ℱt−1F_t-1 the complete history prior to iteration t. The following properties hold. ▶ Exact Anytime Cell Uniformity. For every t≥1t≥ 1 and j∈[M]j∈[M], the spatial allocation of the emitted persona is strictly uniform over the currently active set Pr⁡(c⁡(Ptemit)=j∣ℱt−1)=j∈ℐt−1|ℐt−1|. \! (c(P_t^emit)=j _t-1 )= 1\j _t-1\|I_t-1|. ▶ Finite-Time Discovery. For every T≥1T≥ 1, the probability that the active set has not yet expanded to cover the entire reachable set is bounded by Pr⁡(ℐT≠ℛ)≤∑j∈ℛ∖ℐ0(1−ω0Bj)T. (I_T )≤ _j _0(1- _0B_j)^T. Consequently, all reachable cells are discovered in finite time almost surely. ▶ Asymptotic Convergence. As t→∞t→∞, the law of the emitted persona converges in total variation to the equal-cell mixture ‖ℒ⁡(Ptemit)−Πℛ‖TV⟶0,Πℛ=1|ℛ|∑j∈ℛπj. \|L(P_t^emit)- _R \|_TV 0, _R= 1|R| _j _j. 3.4 Evolutionary TextGrad Persona Generation To actively push personas toward the extreme, unexplored regions of the semantic space, we formulate generation as an evolutionary constrained optimization problem. Unlike UC-MCMC’s distributional target, this is an empirical search that iteratively expands the population tP_t toward valid, low-density semantic frontiers. We evaluate each persona across two geometric axes: Isolation (gap to its nearest neighbor) and Sparsity (unnormalized von Mises-Fisher kernel-density score). Since personas are discrete text rather than continuous vectors, we employ TextGrad (Yuksekgonul et al., 2025) to guide offspring generation with natural-language feedback derived from the geometric fitness scores. At each iteration, the algorithm executes three phases: 1. Parent Selection. A tournament selects elite incumbents as parents that maximize isolation or minimize density. 2. Mutation and Admission. The frozen LLM mutates the parents using the textual gradient on the composite fitness function. Then the offspring is admitted according to the same fitness function and thresholds after the validity checks. 3. Sibling Deduplication. Admitted offspring in the same iteration are mutually deduplicated to remove near-duplicate siblings. Population Final Subset Category Principle (method) Cell- Cov. Hull Cov. @ r0⋆r_0 Disp. Reference PersonaMem (Random) 62.3%62.3\% 4.454.45 0.5890.589 13.6713.67 Select Space filling (Coverage) 62.3%62.3\% 4.454.45 0.9010.901 10.5210.52 Frontier (Dispersion) 62.3%62.3\% 4.454.45 0.0050.005 20.9820.98 Generate Space filling (MCMC) 96.5%96.5\% 6.536.53 0.9050.905 7.917.91 Frontier (Evolution) 76.8%76.8\% 9.479.47 0.0020.002 24.4224.42 Table 1: Persona geometry across persona diversification methods. Population metrics evaluate the candidate pool: cell coverage, which is the occupancy of a fixed 1,0241,024-cell grid, and Mahalanobis hull extent. Subset metrics evaluate the extracted personas used for response generation: Cov.@r0⋆r_0 , which is the coverage objective (eq. 7) with r0⋆r_0 defined as eq. 15 (c=0.9c=0.9 of reference population 0P_0), and dispersion, which is the minimum pairwise Mahalanobis distance. Selection methods share the fixed base pool, while generation methods expand it. Figure 3: 2D projection of persona embeddings illustrating evolutionary frontier generation. The baseline pool (grey) is densely clustered in the center (blue shading), while evolved offspring (orange) populate low-density regions. The downstream selector extracts a k=5k=5 subset (large outlined dots). 4 Evaluation We evaluate our framework on the Alternative Uses Task (AUT), Infinity-Chat (IC), and the Divergent Association Task (DAT). Responses are generated by Gemma-4-31B-it, with semantic embeddings provided by EmbeddingGemma and automated judgments by Qwen3.6-27B. To evaluate transfer without task-specific persona optimization, we hold each task-agnostic persona set fixed across benchmarks. Original task-only prompts are denoted as Alternative-Use on AUT and Baseline Task on IC and DAT. For persona conditioning, Random samples uniformly from the PersonaMem-v2 (Jiang et al., 2025) pool (itself a random subset of PersonaHub (Ge et al., 2024)), while Coverage and Dispersion optimize selections from that pool. MCMC and Evolution expand the candidate pool before applying Coverage and Dispersion selectors, respectively. Prompt baselines include Creativity-enhanced (Góes et al., 2023) on AUT, Creative (Schapiro et al., 2026) on DAT, and reasoning scaffold of DMAD (Diverse Multi-Agent Debate) (Liu et al., 2025). More prompt baselines are compared in Appendix E, but three top-performing baselines are reproduced in Section 4. Combined conditions add the selected personas with the corresponding prompt or reasoning scaffold. Before filtering for validity, each condition generates 25 uses per AUT object, 50 responses per IC query, and 35 ten-noun lists for the DAT. We assess response diversity and creativity alongside validity and utility to identify tradeoffs. We employ a suite of metrics spanning human-derived categorical clustering, LLM-as-a-judge ratings, and semantic embedding metrics. Full prompts, additional baselines and controls, metric definitions, and results are provided in Appendix D–G. RQ1: What persona geometries do the two diversity principles produce at the selection and generation levels? As detailed in Table 1, the framework effectively operationalizes its underlying principles, highlighting a fundamental tension between representation and separation. To contextualize these structural gains, we benchmark against Random as a standard, geometrically blind persona-conditioning pipeline. At the selection level, Coverage covers 90.1%90.1\% of the baseline population, whereas Dispersion increases minimum pairwise distance by 53.5%53.5\% over Random (13.67→20.9813.67→ 20.98). At the generation level, our methods break the fixed pool’s structural boundaries. MCMC expands the population’s cell coverage from 62.3%62.3\% to 96.5%96.5\%, and its Coverage-selected subset covers 90.5%90.5\% of the population at r⋆r . In contrast, Evolution more than doubles the pool’s Mahalanobis hull extent from 4.454.45 to 9.479.47. Figure 3 visualizes this expansion, where the baseline pool remains densely clustered in the center and the evolved offspring deliberately push outward to establish a novel, low-density semantic boundary. Applying Dispersion to this expanded pool yields the highest minimum pairwise distance of 24.4224.42, a 78.6%78.6\% increase over Random. Pool expansion alone can increase the metrics in Table 1 including, coverage, hull extend and dispersion. RQ1 therefore emphasizes the contrastic geometries achieves by two principles: MCMC achives greater cell coverage, whereas Evolution yields greater hull extent despite smaller pool than MCMC. These geometric gains do not guarantee improved response creativity. Therefore, RQ2–RQ4 evaluate whether persona sets optimized without downstream task induce more diverse and creative responses. RQ2: How does persona-set dispersion relate to downstream response diversity? As shown in Figure 4, across 14 configurations per benchmark, greater realized persona-set separation positively correlates with greater mean response dispersion on AUT (r=.40r=.40, descriptive 90% CI [−.19,.78][-.19,.78]), and Infinity-Chat (r=.55r=.55, [.12,.81][.12,.81]). Despite the variance, intervention-aligned comparisons reveal a robust directional trend: all six matched Coverage→ escalations increased task-averaged response dispersion across all 12 contrasts. Additional analyses using Vendi score and convex-hull, showing even stronger correlations between persona-set dispersion and response diversity, appear in Appendix E.2. Figure 4: Persona set and response dispersion on AUT (left) and Infinity-Chat (right). Every points represent different configurations. Arrows connect six matched Coverage→ contrasts while holding candidate-pool and distance metric fixed. Dashed lines show the linear fits and annotations report configuration-level Pearson correlations and 90% confidence intervals. RQ3a: Do diversified task-agnostic personas improve meaningful creativity, and at what cost? Yes. Task-agnostic diversification yields substantial creativity gains with minimal trade-offs. As shown in the top half of Table 2, Random persona conditioning improves diversity by 8.7%8.7\% (1.42→1.541.42→ 1.54), originality by 10.2%10.2\% (10.79→11.8910.79→ 11.89) and creativity by 8.9%8.9\% (OPEN2.72→2.96)2.72→ 2.96) while sacrificing virtually no utility or validity. Our Dispersion selector further improves diversity, originality and flexibility, while our evolutionarily generated personas achieve the highest values across all four target metrics among the standard-prompt conditions. Relative to Alternative-Use, Evolution improves diversity by 78.8%78.8\% (1.42→2.531.42→ 2.53), originality by 26.1%26.1\% (10.79→13.6110.79→ 13.61), flexibility by 49.5%49.5\% (.288→.431.288→.431) and creativity score by 13.9%13.9\% (2.72→3.102.72→ 3.10). Compared to Random, Evolution yields relative improvements ranging from 4.7%4.7\% in creativity up to 64.3%64.3\% in diversity. The cost of these leaps is modest: utility is reduced by just 0.220.22 (−5.1%-5.1\%) from the original task (Alternative-Use), while still producing valid responses 98.5%98.5\% of the time. RQ3b: Do the creativity gains compose with prompt engineering? Yes. Persona diversification composes with, rather than competes against, advanced prompt engineering. As demonstrated in the bottom half of Table 2, applying our evolutionarily generated personas to the creativity-optimized prompt of Góes et al. (2023) amplifies its effects, yielding improvements of +18.6%+18.6\% in diversity, +2.9%+2.9\% in originality, +15.3%+15.3\% in flexibility, and +6.3%+6.3\% in creativity, all while maintaining utility (3.383.38) and improving validity (.968→.979.968→.979). It also shows competitive results against advanced reasoning frameworks, including DMAD (Liu et al., 2025), a state-of-the-art baseline that diversifies reasoning paths rather than personas. Because the personas are generated without optimizing downstream task performance, they are particularly valuable for everyday, zero-shot user queries where elaborate prompt engineering is impractical. Table 2: Response creativity evaluation on AUT. Diversity is the hull extent in a 5-D PCA projection of Mahalanobis-whitened response embeddings, Originality is the mean Mahalanobis distance to the common-use reference set, Flexibility is the number of distinct categories per valid use, and Creativity is the rank-normalized holistic score (Góes et al., 2023) from an independent LLM judge. Subscripts give 95% confidence-interval half-widths across five random seeds. Utility (Stevenson et al., 2022) and validity are controls and are not bolded. Bold = best results within each group. Variant Diversity Originality Flexibility Creativity Utility Validity Alternative-Use 1.42±0.091.42_± 0.09 10.79±0.3210.79_± 0.32 .288±.022.288_±.022 2.72±0.092.72_± 0.09 4.30±0.054.30_± 0.05 .984±.014.984_±.014 Random 1.54±0.121.54_± 0.12 11.89±0.4811.89_± 0.48 .331±.024.331_±.024 2.96±0.042.96_± 0.04 4.28±0.024.28_± 0.02 .993±.006.993_±.006 Dispersion (selection) 1.96±0.171.96_± 0.17 12.00±0.2812.00_± 0.28 .357±.048.357_±.048 2.92±0.042.92_± 0.04 4.23±0.094.23_± 0.09 .986±.010.986_±.010 Evolution (generation) 2.53±0.202.53_± 0.20 13.61±0.1713.61_± 0.17 .431±.045.431_±.045 3.10±0.033.10_± 0.03 4.08±0.064.08_± 0.06 .985±.012.985_±.012 Creativity-enhanced 5.15±0.165.15_± 0.16 17.15±0.4017.15_± 0.40 .503±.056.503_±.056 3.19±0.093.19_± 0.09 3.38±0.073.38_± 0.07 .968±.011.968_±.011 DMAD 2.79±0.222.79_± 0.22 13.90±0.4913.90_± 0.49 .372±.025.372_±.025 2.89±0.072.89_± 0.07 4.17±0.114.17_± 0.11 .977±.009.977_±.009 Evolution + creativity 6.11±0.226.11_± 0.22 17.65±0.1417.65_± 0.14 .580±.046.580_±.046 3.39±0.073.39_± 0.07 3.38±0.163.38_± 0.16 .979±.020.979_±.020 RQ4: Do the task-agnostic gains transfer across tasks and compose with prompt engineering? Yes. Task-agnostic persona diversification transfers robustly across both open-ended and constrained task formats. On Infinity-Chat, Dispersion improves upon Random in all three diversity measures, while Evolution extends these gains further (Table 3). Relative to standard task prompt, Evolution reduces homogeneity (0.946→0.7980.946→ 0.798) and more than doubles flexibility (1.39→2.811.39→ 2.81). Its between-persona separation gain also rises from 0.0680.068 of Random to 0.1260.126, indicating significantly more distinct responses across personas relative to variation within a single persona. Similarly, DAT exhibits the same ordering of escalation across the task-only Baseline Task, Random, Dispersion and Evolution conditions for both divergence and flexibility. On DAT, task-agnostic Evolution also exceeds Creative on these measures. Crucially, despite improved creativity, Evolution maintains near-perfect validity on both benchmarks (99.2%99.2\% on Infinity-Chat and 99.6%99.6\% on DAT), proving that task adherence is not sacrificed. Furthermore, these task-agnostic gains compose seamlessly with advanced reasoning scaffolds. Adding Evolution personas to DMAD further reduces Infinity-Chat homogeneity (0.929→0.9000.929→ 0.900) and raises both Infinity-Chat flexibility (1.56→1.811.56→ 1.81) and DAT score (90.62→90.9390.62→ 90.93), while maintaining validity at 100%100\% and 99.7%99.7\%, respectively, demonstrating the robust compositional benefits. Table 3: Response creativity evaluation on Infinity-Chat (IC) and DAT. Homogeneity is mean pairwise cosine similarity, Separation is mean between-persona distance minus mean within-persona distance, Flexibility is the Vendi score, measuring effective diversity under the embedding similarity kernel, and Divergence is the DAT semantic score. Subscripts give 95% confidence-interval half-widths across five random seeds. Bold = best results within each group. Control (validity) and undefined combinations (–) are not bolded. Infinity-Chat DAT Variant Homog. ↓ Sep. ↑ Flex. ↑ Val. Div. ↑ Flex. ↑ Val. Baseline Task 0.946±0.0020.946_± 0.002 – 1.39±0.021.39_± 0.02 .950±.015.950_±.015 89.13±0.4989.13_± 0.49 1.79±0.011.79_± 0.01 .996±.005.996_±.005 Random 0.864±0.0050.864_± 0.005 0.068±0.0030.068_± 0.003 2.17±0.052.17_± 0.05 1.000±.0001.000_±.000 89.15±1.2489.15_± 1.24 1.83±0.041.83_± 0.04 .993±.011.993_±.011 Dispersion (selection) 0.840±0.0040.840_± 0.004 0.085±0.0060.085_± 0.006 2.40±0.052.40_± 0.05 .986±.007.986_±.007 89.39±0.2689.39_± 0.26 1.84±0.011.84_± 0.01 .995±.006.995_±.006 Evolution (generation) 0.798±0.0060.798_± 0.006 0.126±0.0060.126_± 0.006 2.81±0.082.81_± 0.08 .992±.005.992_±.005 89.48±0.3289.48_± 0.32 1.85±0.011.85_± 0.01 .996±.004.996_±.004 Creative – – – – 86.05±0.4386.05_± 0.43 1.69±0.011.69_± 0.01 .977±.010.977_±.010 DMAD 0.929±0.0020.929_± 0.002 – 1.56±0.011.56_± 0.01 1.000±.0001.000_±.000 90.62±0.2590.62_± 0.25 1.89±0.011.89_± 0.01 .994±.006.994_±.006 Evolution + DMAD 0.900±0.0010.900_± 0.001 0.021±0.0030.021_± 0.003 1.81±0.011.81_± 0.01 1.000±.0001.000_±.000 90.93±0.2990.93_± 0.29 1.89±0.011.89_± 0.01 .997±.008.997_±.008 5 Conclusion This work framed persona-diversity induction as a set-level design problem, distinguishing space-filling from frontier-seeking diversity and selection from generation. The resulting framework provides principled ways to control persona-set geometry and shows that representational diversity and ideational diversity are not the same objective: the persona set that best represents a population need not be the one that elicits the widest range of ideas. More broadly, persona geometry offers a task-agnostic mechanism for shaping the range of LLM outputs, complementary to prompt optimization and inference-time scaffolding, and points toward systems that can deliberately control not only the quality of an answer, but the diversity of perspectives from which answers are generated. AI use statement In this work, we used generative AI tools to help develop theoretical models or conceptual frameworks, formulate mathematical claims, provide critical ingredients for proving mathematical claims, assist in the writing of proofs, design or provide feedback on research methodology or experiments, implement methods, support qualitative and thematic data analysis, and interpret results. We have not used generative AI tools to propose or refine hypotheses and generate synthetic data sets, assist with translation, and clean and reformat dataset are not applicable to this work. Additionally, we used generative AI tools to create or modify scientific figures or images, create or edit software code, creation of artifacts, summarize or analyse existing literature, discover research topics or identify gaps, brainstorming, sourcing/searching for information, edit a research paper to improve readability, identify relevant literature, format references, and propose a title or keywords for a research paper. We have reviewed all AI-assisted work. We reviewed and tested AI-generated experiment code to run as intended, and all conversations with AI are verified before applied to the paper. We take responsibility for the final content of this work, including text, claims or artifacts produced with the aid of generative AI. Reproducibility Statement Sections 2 and 3 define the persona representations, diversification objectives, and algorithms. Appendices B and C provide detailed algorithmic procedures, the assumptions underlying the theoretical guarantees, and their proofs. Appendix D documents the benchmark data sources, baseline construction, generation and preprocessing protocols, evaluation metrics, annotation procedures, and LLM-judge validation. Appendix E provides the full experimental comparisons, with performance estimates and confidence intervals across five generation seeds. Extensive examples of representative personas and prompt templates for persona construction, response generation, and evaluation are provided in Appendices F and G, respectively. References Anthis et al. (2025) Jacy Reese Anthis, Ryan Liu, Sean M. Richardson, Austin C. Kozlowski, Bernard Koch, James Evans, Erik Brynjolfsson, and Michael Bernstein. Llm social simulations are a promising research method, 2025. URL https://arxiv.org/abs/2504.02234. Atmakuru et al. (2024) Anirudh Atmakuru, Jatin Nainani, Rohith Siddhartha Reddy Bheemreddy, Anirudh Lakkaraju, Zonghai Yao, Hamed Zamani, and Haw-Shiuan Chang. Cs4: Measuring the creativity of large language models automatically by controlling the number of story-writing constraints, 2024. URL https://arxiv.org/abs/2410.04197. Azuma (1967) Kazuoki Azuma. Weighted sums of certain dependent random variables. Tohoku Mathematical Journal, 19(3):357–367, 1967. doi: 10.2748/tmj/1178243286. URL https://doi.org/10.2748/tmj/1178243286. BIS (2024) BIS. Artificial intelligence and the economy: Implications for central banks. Technical report, Bank for International Settlements, June 2024. URL https://w.bis.org/publ/arpdf/ar2024e3.htm. Chapter I of the BIS Annual Economic Report 2024. Chakrabarty et al. (2024) Tuhin Chakrabarty, Philippe Laban, Divyansh Agarwal, Smaranda Muresan, and Chien-Sheng Wu. Art or artifice? large language models and the false promise of creativity. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, p. 1–34, 2024. URL https://arxiv.org/abs/2309.14556. Charikar et al. (2001) Moses Charikar, Samir Khuller, David M. Mount, and Giri Narasimhan. Algorithms for facility location problems with outliers. In Proceedings of the Twelfth Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), p. 642–651, 2001. URL https://dl.acm.org/citation.cfm?id=365411.365555. Chen & Ding (2023) Honghua Chen and Nai Ding. Probing the “creativity” of large language models: can models produce divergent semantic association? In Findings of the Association for Computational Linguistics: EMNLP 2023, p. 12881–12888, 2023. URL https://arxiv.org/abs/2310.11158. Choi & Li (2024) Hyeong Kyu Choi and Yixuan Li. PICLe: Eliciting diverse behaviors from large language models with persona in-context learning. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, p. 8722–8739. PMLR, 2024. URL https://proceedings.mlr.press/v235/choi24e.html. Dathathri et al. (2020) Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu. Plug and play language models: A simple approach to controlled text generation, 2020. URL https://arxiv.org/abs/1912.02164. Della Croce et al. (2009) Federico Della Croce, Andrea Grosso, and Marco Locatelli. A heuristic approach for the max–min diversity problem based on max-clique. Computers & Operations Research, 36(8):2429–2433, 2009. doi: 10.1016/j.cor.2008.09.007. URL https://doi.org/10.1016/j.cor.2008.09.007. Erkut (1990) Erhan Erkut. The discrete p-dispersion problem. European Journal of Operational Research, 46(1):48–60, 1990. URL https://doi.org/10.1016/0377-2217(90)90297-O. Feige (1998) Uriel Feige. A threshold of ln n for approximating set cover. Journal of the ACM (JACM), 45(4):634–652, 1998. URL https://doi.org/10.1145/285055.285059. Fernando et al. (2023) Chrisantha Fernando, Dylan Banarse, Henryk Michalewski, Simon Osindero, and Tim Rocktäschel. Promptbreeder: Self-referential self-improvement via prompt evolution, 2023. URL https://arxiv.org/abs/2309.16797. Fillmore (1968) Charles J. Fillmore. The Case for Case. In Emmon Bach and Robert T. Harms (eds.), Universals in Linguistic Theory, p. 1–88. Holt, Rinehart and Winston, New York, NY, 1968. URL https://verbs.colorado.edu/~mpalmer/Ling7800/Fillmore.Case.pdf. Friedman & Dieng (2023) Dan Friedman and Adji Bousso Dieng. The Vendi Score: A diversity evaluation metric for machine learning. Transactions on Machine Learning Research, 2023. URL https://openreview.net/forum?id=g97OHbQyk1. Ge et al. (2024) Tao Ge, Xin Chan, Xiaoyang Wang, Dian Yu, Haitao Mi, and Dong Yu. Scaling synthetic data creation with 1,000,000,000 personas, 2024. URL https://arxiv.org/abs/2406.20094. Gibson (1979) James J. Gibson. The Ecological Approach to Visual Perception. Houghton Mifflin, Boston, MA, 1979. URL https://openlibrary.org/books/OL21229791M/The_Ecological_approach_to_visual_perception. Góes et al. (2023) Luis Fabricio Góes, Marco Volpe, Piotr Sawicki, Marek Grses, and Jacob Watson. Pushing gpt’s creativity to its limits: Alternative uses and torrance tests. International Conference on Computational Creativity, 2023. URL https://computationalcreativity.net/iccc23/papers/ICCC-2023_paper_90.pdf. Guilford (1967) Joy Paul Guilford. The Nature of Human Intelligence. McGraw-Hill, New York, 1967. URL https://search.worldcat.org/title/The-nature-of-human-intelligence/oclc/204270. Hastings (1970) W. K. Hastings. Monte carlo sampling methods using markov chains and their applications. Biometrika, 57(1):97–109, 1970. doi: 10.1093/biomet/57.1.97. URL https://doi.org/10.1093/biomet/57.1.97. Hochbaum & Shmoys (1986) Dorit S. Hochbaum and David B. Shmoys. A unified approach to approximation algorithms for bottleneck problems. Journal of the ACM, 33(3):533–550, 1986. doi: 10.1145/5925.5933. URL https://doi.org/10.1145/5925.5933. Hou et al. (2026) Zhaoyi Joey Hou, Bowei Alvin Zhang, Yining Lu, Bhiman Kumar Baghel, Anneliese Brei, Ximing Lu, Meng Jiang, Faeze Brahman, Snigdha Chaturvedi, Haw-Shiuan Chang, Daniel Khashabi, and Xiang Lorraine Li. Creativityprism: A holistic evaluation framework for large language model creativity, 2026. URL https://arxiv.org/abs/2510.20091. Howard et al. (2021) Steven R. Howard, Aaditya Ramdas, Jon McAuliffe, and Jasjeet Sekhon. Time-uniform, nonparametric, nonasymptotic confidence sequences. The Annals of Statistics, 49(2):1055–1080, 2021. doi: 10.1214/20-AOS1991. URL https://doi.org/10.1214/20-AOS1991. IMF (2024) IMF. Advances in artificial intelligence: Implications for capital market activities. Technical report, International Monetary Fund, October 2024. URL https://w.imf.org/en/Publications/GFSR/Issues/2024/10/22/global-financial-stability-report-october-2024. Chapter 3 of the Global Financial Stability Report: Steadying the Course—Uncertainty, Artificial Intelligence, and Financial Stability. Jiang et al. (2025) Bowen Jiang, Zhuoqun Hao, Young-Min Cho, Bryan Li, Yuan Yuan, Sihao Chen, Lyle Ungar, Camillo J. Taylor, and Dan Roth. Know me, respond to me: Benchmarking llms for dynamic user profiling and personalized responses at scale, 2025. URL https://arxiv.org/abs/2504.14225. Jiang et al. (2026) Liwei Jiang, Yuanjun Chai, Margaret Li, Mickel Liu, Raymond Fok, Nouha Dziri, Yulia Tsvetkov, Maarten Sap, and Yejin Choi. Artificial hivemind: The open-ended homogeneity of language models (and beyond). Advances in Neural Information Processing Systems, 38, 2026. URL https://arxiv.org/abs/2510.22954. Jin et al. (2025) Bojun Jin, Jianzhu Bao, Yufang Hou, Yang Sun, Yice Zhang, Huajie Wang, Bin Liang, and Ruifeng Xu. A multi-persona framework for argument quality assessment. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 12148–12170, 2025. URL https://aclanthology.org/2025.acl-long.593/. Johnson et al. (2013) Alicia A. Johnson, Galin L. Jones, and Ronald C. Neath. Component-wise markov chain monte carlo: Uniform and geometric ergodicity under mixing and composition. Statistical Science, 28(3):360–375, 2013. doi: 10.1214/13-STS423. URL https://doi.org/10.1214/13-STS423. Kirk et al. (2024) Robert Kirk, Ishita Mediratta, Christoforos Nalmpantis, Jelena Luketina, Eric Hambro, Edward Grefenstette, and Roberta Raileanu. Understanding the effects of rlhf on llm generalisation and diversity, 2024. URL https://arxiv.org/abs/2310.06452. Kojima et al. (2022) Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199–22213, 2022. URL https://proceedings.neurips.c/paper/2022/hash/8b0d291acd4acf06ef112099c16f326-Abstract-Conference.html. Kulesza & Taskar (2012) Alex Kulesza and Ben Taskar. Learning determinantal point processes, 2012. URL https://arxiv.org/abs/1202.3738. Kusupati et al. (2022) Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, et al. Matryoshka representation learning. Advances in Neural Information Processing Systems, 35:30233–30249, 2022. URL https://proceedings.nips.c/paper_files/paper/2022/hash/c32319f4868da7613d78af9993100e42-Abstract-Conference.html. Leopardi (2006) Paul Leopardi. A partition of the unit sphere into regions of equal area and small diameter. Electronic Transactions on Numerical Analysis, 25:309–327, 2006. URL https://eudml.org/doc/129860. Levin (1993) Beth Levin. English Verb Classes and Alternations: A Preliminary Investigation. University of Chicago Press, Chicago, IL, 1993. URL https://press.uchicago.edu/ucp/books/book/chicago/E/bo3684144.html. Li et al. (2026) Leon Li, Haozhe Chen, Hongseok Namkoong, and Tianyi Peng. Llm generated persona is a promise with a catch. Advances in Neural Information Processing Systems, 38, 2026. URL https://arxiv.org/abs/2503.16527. Liu et al. (2025) Yexiang Liu, Jie Cao, Zekun Li, Ran He, and Tieniu Tan. Breaking mental set to improve reasoning through diverse multi-agent debate. In The Thirteenth International Conference on Learning Representations, 2025. URL https://proceedings.iclr.c/paper_files/paper/2025/hash/3de667dab3b3d812583abc0a786139a0-Abstract-Conference.html. Lu et al. (2025) Ximing Lu, Melanie Sclar, Skyler Hallinan, Niloofar Mireshghallah, Jiacheng Liu, Seungju Han, Allyson Ettinger, Liwei Jiang, Khyathi Chandu, Nouha Dziri, et al. Ai as humanity’s salieri: Quantifying linguistic creativity of language models via systematic attribution of machine text against web text. In International Conference on Learning Representations, p. 91005–91056, 2025. URL https://arxiv.org/abs/2410.04265. Mengersen & Tweedie (1996) Kerrie L. Mengersen and Richard L. Tweedie. Rates of convergence of the hastings and metropolis algorithms. The Annals of Statistics, 24(1):101–121, 1996. doi: 10.1214/aos/1033066201. URL https://doi.org/10.1214/aos/1033066201. Mouret & Clune (2015) Jean-Baptiste Mouret and Jeff Clune. Illuminating search spaces by mapping elites, 2015. URL https://arxiv.org/abs/1504.04909. Nemhauser et al. (1978) George L Nemhauser, Laurence A Wolsey, and Marshall L Fisher. An analysis of approximations for maximizing submodular set functions—i. Mathematical programming, 14(1):265–294, 1978. ISSN 1436-4646. doi: 10.1007/BF01588971. URL https://doi.org/10.1007/BF01588971. Nguyen & Dieng (2024) Quan Nguyen and Adji Bousso Dieng. Quality-weighted vendi scores and their application to diverse experimental design. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, p. 37667–37682. PMLR, 2024. URL https://proceedings.mlr.press/v235/nguyen24d.html. Norman (1988) Donald A. Norman. The Psychology of Everyday Things. Basic Books, New York, NY, 1988. URL https://openlibrary.org/books/OL2409159M/The_psychology_of_everyday_things. Olson et al. (2021) Jay A Olson, Johnny Nahas, Denis Chmoulevitch, Simon J Cropper, and Margaret E Webb. Naming unrelated words predicts creativity. Proceedings of the National Academy of Sciences, 118(25):e2022340118, 2021. URL https://w.pnas.org/doi/10.1073/pnas.2022340118. Paglieri et al. (2026) Davide Paglieri, Logan Cross, William A. Cunningham, Joel Z. Leibo, and Alexander Sasha Vezhnevets. Persona generators: Generating diverse synthetic personas for arbitrary contexts, 2026. URL https://arxiv.org/abs/2602.03545. Pasarkar & Dieng (2024) Amey P Pasarkar and Adji Bousso Dieng. Cousins of the Vendi Score: A family of similarity-based diversity metrics for science and machine learning. In International Conference on Artificial Intelligence and Statistics (AISTATS), 2024. URL https://proceedings.mlr.press/v238/pasarkar24a.html. Rabeyah et al. (2025) Abdullah Al Rabeyah, Fabrício Góes, Marco Volpe, and Talles Medeiros. Do llms agree on the creativity evaluation of alternative uses? International Conference on Computational Creativity, 2025. URL https://computationalcreativity.net/iccc25/papers/iccc25-rabeyah2025do.pdf. Roberts & Rosenthal (2007) Gareth O. Roberts and Jeffrey S. Rosenthal. Coupling and ergodicity of adaptive Markov chain Monte Carlo algorithms. Journal of Applied Probability, 44(2):458–475, 2007. doi: 10.1239/jap/1183667414. URL https://doi.org/10.1239/jap/1183667414. Rosenthal (1995) Jeffrey S. Rosenthal. Minorization conditions and convergence rates for markov chain monte carlo. Journal of the American Statistical Association, 90(430):558–566, 1995. doi: 10.1080/01621459.1995.10476548. URL https://doi.org/10.1080/01621459.1995.10476548. Runco & Jaeger (2012) Mark A Runco and Garrett J Jaeger. The standard definition of creativity. Creativity Research Journal, 24(1):92–96, 2012. URL https://doi.org/10.1080/10400419.2012.650092. Schapiro et al. (2026) Samuel Schapiro, Core Francisco Park, Felix Sosa, and Lav R Varshney. Creativityneuro: Steering language model weights to improve divergent thinking and reduce mode collapse. arXiv preprint arXiv:2607.01433, 2026. URL https://arxiv.org/abs/2607.01433. Sorensen et al. (2024) Taylor Sorensen, Jared Moore, Jillian Fisher, Mitchell Gordon, Niloofar Mireshghallah, Christopher Michael Rytting, Andre Ye, Liwei Jiang, Ximing Lu, Nouha Dziri, Tim Althoff, and Yejin Choi. A roadmap to pluralistic alignment, 2024. URL https://arxiv.org/abs/2402.05070. Sorensen et al. (2026) Taylor Sorensen, Benjamin Newman, Jared Moore, Chan Park, Jillian Fisher, Niloofar Mireshghallah, Liwei Jiang, and Yejin Choi. Spectrum tuning: Post-training for distributional coverage and in-context steerability, 2026. URL https://arxiv.org/abs/2510.06084. Stevenson et al. (2022) Claire Stevenson, Iris Smal, Matthijs Baas, Raoul Grasman, and Han van der Maas. Putting gpt-3’s creativity to the (alternative uses) test. International Conference on Computational Creativity, 2022. URL https://computationalcreativity.net/iccc22/wp-content/uploads/2022/06/ICCC-2022_25S_Stevenson-et-al..pdf. Su & Collier (2023) Yixuan Su and Nigel Collier. Contrastive search is what you need for neural text generation, 2023. URL https://arxiv.org/abs/2210.14140. Tierney (1994) Luke Tierney. Markov chains for exploring posterior distributions. The Annals of Statistics, 22(4):1701–1762, 1994. doi: 10.1214/aos/1176325750. URL https://doi.org/10.1214/aos/1176325750. Wei et al. (2015) Kai Wei, Rishabh Iyer, and Jeff Bilmes. Submodularity in data subset selection and active learning. In Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, p. 1954–1963. PMLR, 2015. URL https://proceedings.mlr.press/v37/wei15.html. West & Potts (2025) Peter West and Christopher Potts. Base models beat aligned models at randomness and creativity, 2025. URL https://arxiv.org/abs/2505.00047. Yuksekgonul et al. (2025) Mert Yuksekgonul, Federico Bianchi, Joseph Boen, Sheng Liu, Pan Lu, Zhi Huang, Carlos Guestrin, and James Zou. Optimizing generative ai by backpropagating language model feedback. Nature, 639:609–616, 2025. doi: 10.1038/s41586-025-08661-4. URL https://doi.org/10.1038/s41586-025-08661-4. Zheng et al. (2024) Huaixiu Steven Zheng, Swaroop Mishra, Xinyun Chen, Heng-Tze Cheng, Ed H. Chi, Quoc V Le, and Denny Zhou. Take a step back: Evoking reasoning via abstraction in large language models. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=3bq3jsvcQ1. Appendix Contents Appendix A Related Works Persona-conditioned and personalized language modeling. Role prompting is a widely used interface for steering large language models, but recent work suggests that a persona can be treated as a more structured object than a generic expert role. PersonaHub scales persona-conditioned data synthesis to a billion synthetic personas, showing that persona descriptions can serve as reusable carriers of perspective and task variation (Ge et al., 2024). In parallel, work on personalized response generation has emphasized that user profiles are dynamic, sparse, and context-dependent: PersonaMem benchmarks LLMs on maintaining and using evolving user profiles over long interaction histories (Jiang et al., 2025). These efforts establish personas as useful conditioning variables, but they primarily focus on sampling, profiling, or evaluating personalization. Other work cautions that LLM-generated personas may encode systematic biases or brittle assumptions, especially when used for downstream simulation or evaluation (Li et al., 2026). In contrast, our work studies personas as optimizable latent controls: rather than relying on generic roles or fixed persona pools, we extract, adapt, select, and contrast personas to improve the novelty and diversity of model generations. Diversity-aware generation and optimization. A long line of work has studied diversity in generation and subset selection. Determinantal point processes provide a principled probabilistic model for selecting diverse subsets while balancing quality and diversity (Kulesza & Taskar, 2012). Submodular facility-location objectives similarly formalize coverage and redundancy reduction, with classical greedy algorithms providing approximation guarantees for monotone submodular maximization under cardinality constraints (Nemhauser et al., 1978). In evolutionary computation, quality-diversity methods such as MAP-Elites search for collections of high-performing but behaviorally distinct solutions, rather than optimizing a single best point (Mouret & Clune, 2015). Recent LLM work has brought related ideas into prompt and decoding optimization: Promptbreeder evolves task prompts through population-based self-improvement (Fernando et al., 2023), Plug-and-Play Language Models steer generation through gradient-based control without retraining the base model (Dathathri et al., 2020), and Contrastive Search improves open-ended generation by discouraging degeneration while preserving coherence (Su & Collier, 2023). Our approach builds on these ideas but changes the optimization target: we optimize persona populations rather than only prompts, logits, or samples, with the goal of producing useful creative diversity rather than only improving task accuracy or surface-level lexical variation. Pluralism, coverage, and model homogeneity. The motivation for optimizing persona diversity is closely related to recent concerns about output-space collapse. Work on pluralistic alignment argues that many tasks admit a spectrum of valid responses, and that alignment should preserve calibrated diversity rather than collapse toward a single normative answer (Sorensen et al., 2024). Spectrum Tuning further operationalizes this idea through distributional coverage and in-context steerability, showing that post-training can affect how well models cover diverse valid responses (Sorensen et al., 2026). Related evidence on artificial hiveminds suggests that LLMs can exhibit substantial homogeneity on open-ended tasks, even when many distinct responses would be reasonable (Jiang et al., 2026). These findings motivate our central hypothesis: increasing structured diversity in the persona or prompt space can increase diversity in the response space, provided that the personas remain task-relevant and high quality. Creativity benchmarks and metrics. Evaluating creativity in LLMs is challenging because novelty, usefulness, surprise, fluency, and diversity can diverge. Classic divergent-thinking paradigms such as the Alternative Uses Test have been adapted to evaluate whether LLMs can produce semantically distant but plausible uses for everyday objects (Stevenson et al., 2022), while the Divergent Association Task probes whether models can generate remote semantic associations (Chen & Ding, 2023). For creative writing, TTCW shows that LLM-generated stories often fall short of professional human writing under expert evaluation (Chakrabarty et al., 2024), and CS4 studies creativity under increasing numbers of story-writing constraints, highlighting trade-offs between originality, coherence, and instruction following (Atmakuru et al., 2024). Complementary metric work such as the Creativity Index quantifies linguistic novelty through attribution against web text (Lu et al., 2025), while CreativityPrism evaluates creativity holistically across tasks, domains, and metrics (Hou et al., 2026). These benchmarks motivate our evaluation design: to show that persona-conditioned generation improves creativity, it is not sufficient to produce stylistically different outputs; the outputs must be measurably more novel, diverse, coherent, and useful across multiple creativity settings. Appendix B Selection Algorithms and Proofs B.1 Adaptive Thresholded Coverage B.1.1 Exact ILP formulation of the coverage objective As outlined in Section 3.2, we formulate persona selection as a maximal-covering location problem rather than an additive maximum-similarity objective. While an additive model maximizes the summed similarity of every persona to its nearest exemplar (Wei et al., 2015), its marginal gain weights every demand point’s similarity improvement, allowing dense regions of the persona space to disproportionately dominate the objective. By contrast, under our thresholded coverage model, a candidate exemplar’s marginal gain is strictly the number of previously uncovered demand personas inside its radius-r neighborhood; already-covered points contribute zero additional gain. (Under cosine dissimilarity, this effectively thresholds the cosine similarity at τcov=1−r _cov=1-r, yielding a τcov _cov-coverage model). Recall the coverage objective for a selected subset S of cardinality k at a fixed radius r: Fcov,m(,,r)F_cov,m(S;P,r) (equation 7). We adopt the convention min⁡∅=+∞ =+∞, ensuring that no persona is covered by the empty set and Fcov,m(∅,,r)=0F_cov,m( ;P,r)=0. At a fixed radius r, the maximum-coverage persona-selection problem seeks a subset cov,m⋆(,r)S _cov,m(P;r) of cardinality k that leaves as little of the candidate population uncovered as possible cov,m⋆(,r)∈arg⁡max⊆||=kFcov,m(,,r).S _cov,m(P;r)∈ _ subarraycS \\ |S|=k subarrayF_cov,m(S;P,r). (9) Because the inner min operator and the indicator function render this objective (equation 7) non-linear, we linearize it exactly. First, we define the coverage neighborhood of persona pip_i, which always contains i itself mcov(i,r)=j∈[N]:dmP(pi,pj)≤r,J^cov_m(i;r)= \j∈[N_P]:d_m^P(p_i,p_j)≤ r \, (10) We then introduce two sets of binary decision variables • χj∈0,1 _j∈\0,1\: selection indicator for exemplar pjp_j; • zicov∈0,1z_i^cov∈\0,1\: covered indicator for candidate persona pip_i. Writing =(χ1,…,χN)⊤ χ=( _1,…, _N_P) and cov=(z1cov,…,zNcov)⊤z^cov=(z_1^cov,…,z_N_P^cov) , our goal is to select exactly k exemplars to maximize the unnormalized covered count. Structuring the objective to be integer-valued allows us to pose the exact maximal covering location model as the following Integer Linear Program (ILP) maximize,cov χ,z^covmaximize ∑i=1Nzicov _i=1^N_Pz_i^cov (11) subject to zicov≤∑j∈mcov(i,r)χj z_i^cov≤ _j ^cov_m(i;r) _j ∀i∈[N], ∀ i∈[N_P], (12) ∑j=1Nχj=k, _j=1^N_P _j=k, (13) χj∈0,1,zicov∈0,1 _j∈\0,1\, z_i^cov∈\0,1\ ∀i,j∈[N]. ∀ i,j∈[N_P]. (14) Dividing the optimum of this solver objective by N_P perfectly recovers Fcov,mF_cov,m (equation 7). Remark 1 (Model Size). This program requires only 2N2N_P binary variables and N+1N_P+1 constraints. This is highly efficient compared to the N2N_P^2 continuous assignment variables and N2N_P^2 linking constraints required by the strong k-median formulation of the additive objective. Furthermore, the dissimilarities enter the coverage constraint (equation 12) only through the Boolean level sets mcov(i,r)J^cov_m(i;r) (equation 10), meaning no similarity normalization, frozen reference scale, or integer rescaling of distances is needed to pose the model. B.1.2 Formulation properties The thresholded coverage objective avoids the complex scale bookkeeping required by additive formulations. Its key structural properties stem from the fact that the optimization depends on the dissimilarity metric strictly through its Boolean level sets. Theorem 4 (Ordinal Invariance). Let g:ℝ≥0→ℝ≥0g:R_≥ 0 _≥ 0 be strictly increasing function with g⁡(0)=0g(0)=0. Replacing the dissimilarity metric dmPd_m^P with g∘dmPg d_m^P and the radius r by g⁡(r)g(r) leaves every coverage neighborhood mcov(i,r)J^cov_m(i;r) (equation 10) unchanged. Consequently, the feasible set, the objective function, and the optimizers of the ILP formulation (equation 11–equation 14) remain strictly identical. Proof. The program depends on the dissimilarity metric solely through the sets mcov(i,r)J^cov_m(i;r). Because g is strictly increasing, the inequality dmP(pi,pj)≤rd_m^P(p_i,p_j)≤ r holds if and only if g⁡(dmP(pi,pj))≤g⁡(r)g(d_m^P(p_i,p_j))≤ g(r). Therefore, every level set, and the entire optimization program by extension, is perfectly preserved. ∎ In practice, this means that formulating the cosine variant as a similarity threshold τcov _cov or as a distance radius r=1−τcovr=1- _cov selects the exact same personas. No frozen reference scale is required to ensure the objective is non-negative or strictly comparable across different candidate populations. Theorem 5 (Integrality of Relaxed Coverage Indicators). If only the covered indicators are relaxed to continuous bounds zicov∈[0,1]z_i^cov∈[0,1], then for any fixed, feasible binary selection vector ∈0,1N χ∈\0,1\^N_P, the relaxed program is maximized by (zicov)⋆=min⁡1,∑j∈mcov(i,r)χj∈0,1.(z_i^cov) = \1, _j ^cov_m(i;r) _j \∈\0,1\. Thus, this relaxation is exact: integrality is inherently guaranteed for the covered-indicator block whenever χ is binary. (Note: This does not assert integrality of the full LP relaxation where χ is also relaxed.) Proof. The objective is non-decreasing in each zicovz_i^cov, and distinct rows are coupled only through χ. Therefore, the solver independently raises each zicovz_i^cov to the largest value permitted by the coverage constraint (equation 12) and the upper bound zicov≤1z_i^cov≤ 1. That value is the minimum of 11 and a non-negative integer, which is inherently binary. ∎ Remark 2 (Demand Aggregation). Personas sharing identical coverage neighborhoods at radius r can be merged into a single demand row with an integer weight without altering the optimum. While near-duplicate personas may share their neighborhoods and can thus be aggregated, spatial proximity alone does not guarantee identical neighborhood rows. Lemma 1 (Monotone Submodularity). The unnormalized covered count ↦NFcov,m(,,r)S N_P\,F_cov,m(S;P,r) is normalized, monotone, and submodular. Therefore, a standard greedy algorithm attains at least a (1−1/e)(1-1/e) fraction of the optimal covered count (Nemhauser et al., 1978). Proof. The covered count evaluates to |⋃pj∈i∈[N]:j∈mcov(i,r)| | _p_j \i∈[N_P]:j ^cov_m(i;r)\ |, which is the cardinality of a union of finite sets indexed by the selected exemplars. This constitutes a classical coverage function, and all coverage functions are normalized, monotone, and submodular. ∎ As previewed in Section 3.2, we utilize Lemma 1 not as an approximation guarantee (since our final selection algorithm below is exact), but as the mathematical engine for the cheap, one-sided greedy certificates that keep the exact solver out of most of the vast majority of the radius search space. B.1.3 Adaptive radius calibration As outlined in Section 3.2, manually fixing the radius r would reintroduce exactly the dissimilarity-specific scaling that Theorem 4 explicitly eliminates. We therefore calibrate it from a scale-free parameter: a target coverage level ccov∈(0,1]c_cov∈(0,1], mapping to an absolute covered-count target ncov=⌈ccovN⌉n_cov= c_covN_P . The calibrated radius is defined as the smallest possible radius at which some size-k subset successfully meets the target rm⋆(,ccov)=min⁡r≥0:max⊆||=k⁡NFcov,m(,,r)≥ncov,r _m(P;c_cov)= \r≥ 0: _ subarraycS \\ |S|=k subarrayN_P\,F_cov,m(S;P,r)≥ n_cov \, (15) In a metric space, this formulation specialized to the discrete generalized k-center-with-outliers objective (Charikar et al., 2001), accommodating N−ncovN_P-n_cov outliers. (Note: Our exact finite search operates strictly on pairwise dissimilarities and does not assume a triangle inequality.) The final selection algorithm is lexicographic: it first discovers the minimal radius rm⋆(,ccov)r _m(P;c_cov) (equation 15), and then break ties among radius-optimal subsets by maximizing the covered count (equation 11) at that exact radius. We denote this calibrated selector by acov,m⋆(,ccov)∈arg⁡max⊆||=kFcov,m(,,rm⋆(,ccov)).S _acov,m(P;c_cov)∈ _ subarraycS \\ |S|=k subarrayF_cov,m\! (S;P,r _m(P;c_cov) ). (16) Lemma 2 (Finite Candidate Radii). For the finite pairwise spectrum Λm() _m(P) defined in equation 6, every fixed subset S has a covered count that is non-decreasing in r and constant between consecutive values of Λm() _m(P). Consequently, the same holds for its maximum over all size-k subsets. The feasible set of the adaptive radius objective (equation 15) is therefore the half-line [rm⋆(,ccov),∞)[r _m(P;c_cov),∞), and the exact minimum rm⋆(,ccov)∈Λm()r _m(P;c_cov)∈ _m(P). Proof. As r grows, the indicator function in the coverage objective (equation 7) can change only when r crosses a realized dissimilarity in the matrix, as the coverage events minpj∈dmP(pi,pj)≤r\ _p_j d_m^P(p_i,p_j)≤ r\ are nested in r. Hence, every covered count is a non-decreasing step function whose jumps lie exactly in Λm() _m(P). Since the pointwise maximum over the finitely many size-k subsets is also a step function with jumps in Λm() _m(P), the smallest feasible radius is inherently attained at a jump point. ∎ By Lemma 2, calibration reduces to an exact discrete search over the sorted distinct values of Λm() _m(P) (Hochbaum & Shmoys, 1986). To avoid invoking the computationally heavy ILP solver for every probe, we evaluate the monotone decision problem of whether k exemplars can cover at least ncovn_cov personas at radius r, using a standard size-k greedy maximum-coverage construction. Let Gm(r)G_m(r) denote the covered count returned by this greedy heuristic. Three certificate mechanisms resolve the vast majority of search probes instantly: 1. Greedy feasibility certificate. If Gm(r)≥ncovG_m(r)≥ n_cov, the probed radius is immediately certified as feasible. 2. Greedy infeasibility certificate. By the monotone submodularity established in Lemma 1, the optimal covered count is strictly bounded above by ⌊Gm(r)/(1−1/e)⌋ G_m(r)/(1-1/e) . If this proven upper bound falls below ncovn_cov, the probed radius is mathematically certified as infeasible. 3. Certificate tightening. Any feasible selection S remains feasible down to its own minimal radius (the ncovn_cov-th smallest nearest-exemplar dissimilarity under S). This allows the feasible bracket to jump far below the initially probed radius in a single computational step. Probes falling into the narrow residual window are resolved via bounded optimization. The exact ILP solver attempts to maximize the covered count but is instructed to terminate early the moment an incumbent reaches ncovn_cov (a feasibility certificate) or its proven upper bound drops below ncovn_cov (an infeasibility certificate that remains mathematically valid even if cut off by a time limit, because the objective is strictly integer-valued). By monotonicity, the search terminates as soon as it certifies that the immediate candidate predecessor of rm⋆(,ccov)r _m(P;c_cov) is infeasible, whenever such a predecessor exists. That final certificate may come from the greedy upper bound or the exact solver; no lower-candidate certificate is needed when the optimum is rm,(1)r_m,(1), the absolute minimum dissimilarity in the spectrum. B.1.4 Exact algorithm Maximum coverage is NP-hard (Feige, 1998) when the selection budget is part of the input, and exact optimization can be computationally expensive in the worst-case. However, our compact ILP formulation (equation 11–equation 14) is highly tractable in practice. To resolve the exact probes during discrete search, the formulation is warm-started with the greedy incumbent and evaluated by an exact integer or constraint solver (e.g., CP-SAT). We accept a final solution as exact only when it achieves certified optimal status. Because the unnormalized covered count is integral, an incumbent solution is automatically certified as optimal once its proven mathematical upper bound drops less than one unit above it. Algorithm 1 Exact Adaptive Thresholded-Coverage Selection 1: Input: Candidate population P of size N_P, dissimilarity index m, target size k, target coverage ccovc_cov 2: Embed each pi∈p_i ⊳ ⁡(N)O(N_P) encoder calls 3: Compute Dm,ijP←dmP(pi,pj)D^P_m,ij← d_m^P(p_i,p_j) and mP←(Dm,ijP)i,j=1ND_m^P←(D^P_m,ij)_i,j=1^N_P ⊳ ⁡(N2)O(N_P^2) pair evaluations 4: ncov←⌈ccovN⌉n_cov← c_covN_P ⊳ ⁡(1)O(1) 5: Form Λm() _m(P) and sort it as rm,(1)<⋯<rm,(Lm)r_m,(1)<·s<r_m,(L_m) ⊳ naively ⁡(N2log⁡N)O(N_P^2 N_P) 6: ℓhi←Lm _hi← L_m with any size-k witness feasible at rm,(Lm)r_m,(L_m); ℓlo←0 _lo← 0 ⊳ 00 is a sentinel, never a radius index 7: Apply greedy certificates and witness tightening to the bracket, preserving feasible ℓhi _hi and either ℓlo=0 _lo=0 or certified-infeasible ℓlo _lo 8: while ℓlo<ℓhi−1 _lo< _hi-1 do 9: Probe ℓ←ℓhi−1 ← _hi-1 at radius rm,(ℓ)r_m,( ): greedy certificates, else the bounded-optimization oracle 10: if feasible, with witness cover S then 11: ℓhi← _hi← index of the minimal radius of S ⊳ certificate tightening 12: else 13: ℓlo←ℓ _lo← ⊳ certified infeasible 14: end if 15: end while 16: r⋆←rm,(ℓhi)r ← r_m,( _hi) 17: Solve equation 11–equation 14 at r=r⋆r=r to certified optimality, warm-started with the incumbent cover 18: Output: Globally optimal coverage subset acov,m⋆(,ccov)S _acov,m(P;c_cov) and calibrated radius rm⋆(,ccov)=r⋆r _m(P;c_cov)=r Remark 3 (Probe Schedule and Time Limits). Any probe schedule that terminates with the tight bracket ℓlo=ℓhi−1 _lo= _hi-1 mathematically certifies the same minimal radius. Alternative schedules can change both the number and difficulty of radius decisions. For example, sequential probing can require O⁡(Lm)O(L_m) decisions, whereas binary search requires O⁡(log⁡Lm)O( L_m) certified decisions. If a bounded-optimization probe is cut off by an imposed solver time limit and is therefore conservatively treated as infeasible, the returned selection is still a valid optimal cover at the returned radius. However, that radius acts only as an upper bound on the true minimal radius rm⋆(,ccov)r _m(P;c_cov). The exact optimality guarantee established in Theorem 1 applies only when every decisive infeasibility result is formally certified. B.1.5 Optimality guarantee Proof. The proof of global optimality relies on three sequential guarantees: Termination. The discrete distance spectrum Λm() _m(P) is finite. Since every iteration of the while-loop in Algorithm 1 must either strictly lower the upper bound ℓhi _hi by at least one index or raise the lower bound ℓlo _lo to ℓhi−1 _hi-1, the search space strictly shrinks at every step. Therefore, the algorithm is guaranteed to terminate in finite time. Radius optimality. Two invariants are maintained throughout the execution of the search loop. First, the radius rm,(ℓhi)r_m,( _hi) is always feasible, guaranteed by a stored size-k witness cover. (Certificate tightening preserves this invariant because any cover is naturally feasible down to its own internal minimal radius by construction.) Second, either ℓlo=0 _lo=0 (acting as the initialized sentinel) or rm,(ℓlo)r_m,( _lo) is certified infeasible. This refutation is valid whether it comes from the greedy certificate (validated by the submodularity established in Lemma 1) or from the exact solver (whose proven bound acts as a strict upper bound on the true integral optimum.) Upon termination, the bracket perfectly tightens. Either ℓhi=1 _hi=1, meaning the smallest candidate radius is in the spectrum is feasible, or ℓlo=ℓhi−1≥1 _lo= _hi-1≥ 1, meaning the immediate predecessor to the upper bound is infeasible. By monotonicity and the discrete step-function property established in Lemma 2, this guarantees that the returned radius is the global minimum r⋆=rm,(ℓhi)=rm⋆(,ccov).r =r_m,( _hi)=r _m(P;c_cov). Coverage optimality at r⋆r . The final step of the algorithm executes a bounded optimization over the finite set of selection vectors ∈0,1N χ∈\0,1\^N_P subject to the exact cardinality constraint ∑jχj=k _j _j=k. The ILP solver evaluates this space and yields a feasible incumbent cover alongside a proven upper bound on the maximum possible covered count. Achieving a certified optimal status guarantees that these two values are equal. Furthermore, because the unnormalized covered count is integer-valued, proving an absolute gap of less than one is sufficient to rule out the existence of a better feasible selection. ∎ B.2 Max–Min Dispersion Persona Selection B.2.1 Structural properties and thresholded equivalence Unlike greedy heuristics, the exact incremental-clique method relies on the discrete topology of the solution space rather than the continous geometric properties of the embedding space. Assumption 1 (Pairwise dissimilarity). The chosen dissimilarity function dmP(⋅,⋅)d_m^P(·,·) is symmetric and non-negative. However, a triangle inequality is not required for the exact clique reduction to hold. The theoretical guarantee of this exact search is grounded in the finite, discrete nature of the candidate pool: Lemma 3 (Bottleneck Distance, (Erkut, 1990)). Because the candidate population P is finite, the optimal value Fdisp,m(disp,m⋆())F_disp,m(S _disp,m(P)) equals the weight of at least one edge in the complete pairwise-dissimilarity graph on constructed on P. B.2.2 Exact incremental-clique algorithm Algorithm 2 Exact Incremental-Clique Max–Min Dispersion 1: Input: Candidate population P, target size k, persona dissimilarity dmPd_m^P 2: Compute the distinct pairwise-dissimilarity levels Λm() _m(P) and order them as rm,(1)<⋯<rm,(Lm)r_m,(1)<·s<r_m,(L_m). 3: Initialize =(,ℰ)G=(P,E) with ℰ←∅E← 4: for ℓ=Lm,Lm−1,…,1 =L_m,L_m-1,…,1 do 5: ℰℓ←pa,pb:a<b,dmP(pa,pb)=rm,(ℓ)E_ ←\\p_a,p_b\:a<b,\ d_m^P(p_a,p_b)=r_m,( )\ 6: for each edge pa,pb∈ℰℓ\p_a,p_b\ _ do 7: ℰ←ℰ∪pa,pbE ∪\\p_a,p_b\\ 8: a,b←pc∈:pa,pc∈ℰ∧pb,pc∈ℰN_a,b←\p_c :\p_a,p_c\ \p_b,p_c\ \ 9: if |a,b|≥k−2|N_a,b|≥ k-2 then 10: Test whether ⁡[a,b]G[N_a,b] contains a (k−2)(k-2)-clique 11: if a (k−2)(k-2)-clique Q exists then 12: Return disp,m⋆()←∪pa,pbS _disp,m(P) ∪\p_a,p_b\ 13: end if 14: end if 15: end for 16: end for B.2.3 Optimality guarantee Proof. Let δ⋆δ denote the optimal max–min dispersion. By Lemma 3 (the bottleneck lemma), δ⋆=rm,(ℓ⋆)δ =r_m,( ) for some threshold in the sorted distance spectrum. After all edges at threshold rm,(ℓ)r_m,( ) have been inserted, Algorithm 2 has effectively constructed the threshold graph ℓ=(,p,p′:p≠p′,dmP(p,p′)≥rm,(ℓ)).G_ = (P,\\p,p \:p≠ p ,\ d_m^P(p,p )≥ r_m,( )\ ). A size-k subset has a minimum internal dispersion of at least rm,(ℓ)r_m,( ) if and only if it forms a full k-clique in ℓG_ . The algorithm processes thresholds from largest to smallest and returns as soon as the first k-clique is completed. If a subset with a strictly larger dispersion existed, its corresponding clique would have been fully connected at an earlier (larger) threshold and the algorithm would have already terminated. Hence the returned clique inherently possesses the bottleneck distance rm,(ℓ⋆)=δ⋆r_m,( )=δ , proving it is globally optimal. ∎ Remark 4 (Computational Efficiency via Sparsity). Although finding a size-k clique is NP-hard in general, the threshold graphs evaluated during the earliest stages of the algorithm (the largest distances) are sparse. For our experiment setting of k=5k=5, each local feasibility test reduces to triangle detection in the mutual neighborhood a,bN_a,b. A direct implementation has a coarse polynomial bound of O⁡(N5)O(N_P^5), while this sparsity reduce actual search costs. Appendix C Generation Algorithms and Proofs C.1 Uniform-Coverage MCMC Persona Generation We propose Uniform-Coverage MCMC (UC-MCMC), a persona-generation sampler whose operational target assigns equal probability mass to the reachable cells of a fixed, finite-resolution partition of semantic directions. Within each cell, a frozen language model supplies a reference base law favoring linguistically plausible personas; valid proposals that land in previously unoccupied cells act to expand the active target in hindsight. The construction successfully pursues coverage through its sampling law, rather than by first generating a pool and then retrospectively equalizing its cell counts. C.1.1 Semantic Geometry and Target Distributions State Space and Validity. We specialize Ω to the countable persona state space of canonically serialized, EOS-terminated persona token sequences up to a fixed maximum length. A nonterminating generation is represented by a failure symbol ⊥∉Ω ∉ . Let val:Ω→0,1val: →\0,1\ be a fixed validity function. This function may incorporate a cached LLM judgment, provided that the model, prompt, decoding rule, and cache are frozen so that val⁡(p)val(p) is deterministic. We extend this function with the convention val⁡(⊥)=0val( )=0. Semantic Geometry and Cell Partitions. Let ϕ~ φ be the MRL-truncated text embedding, whose restriction to the persona state space maps Ω to ℝMdR^d_M. We freeze a center μ μ and a nonsingular whitening map W (computed using only the baseline population or a separate pilot sample). The semantic direction of a persona is defined as its normalized projection W(p)=Wϕ~(p)−μ^‖W⁡ϕ~(p)−μ^‖2∈dM−1.z_W(p)= W\ φ(p)- μ\ \|W\ φ(p)- μ\ \|_2 ^d_M-1. (17) We assume the denominator is nonzero for every valid persona. We then partition the sphere dM−1S^d_M-1 into M measurable, equal-area cells 1,…,MC_1,…,C_M, with deterministic tie-breaking on cell boundaries. (Equal-area, small-diameter sphere partitions can be constructed using, for example, the recursive zonal method of Leopardi (2006).) We define the cell-assignment function c⁡(p)=jc(p)=j when W(p)∈jz_W(p) _j, allowing use to formally define the valid persona set within cell j as Ωj=p∈Ω:val(p)=1,c(p)=j. _j=\p∈ :val(p)=1,\ c(p)=j\. (18) Note that equal area is required for the geometric interpretation formalized below, but the underlying MCMC validity results require only a fixed measurable partition. Base Law and Cell-Conditioned Targets. Let GθPG_ _P be the frozen persona-generation model and let xgenx_gen be its fully serialized, fixed prompt. For persona p, let w1:L⁡(p)w_1:L(p) be its exact output token sequence, excluding terminal EOS. We define the base probability mass of p as b⁡(p)=∏ℓ=1L⁡(p)GθP(wℓ∣xgen,w<ℓ)GθP(EOS∣xgen,w≤L⁡(p)).b(p)= \ _ =1^L(p)G_ _P(w_ x_gen,w_< ) \G_ _P(EOS x_gen,w_≤ L(p)). (19) Every generation outcome outside Ω is mapped to the failure symbol ⊥ without resampling. Thus, we set b⁡(⊥)=1−∑p∈Ωb⁡(p)b( )=1- _p∈ b(p) as the probability law on Ω⊥:=Ω∪⊥ _ := ∪\ \, while ⊥ carries zero target weight because val⁡(⊥)=0val( )=0. We assume b⁡(p)>0b(p)>0 for every valid persona, which naturally holds for raw softmax sampling when all canonical output tokens remain in support. In our primary configuration, the global independence proposal is exactly this completion law q0(p′∣p)=q0(p′)=b⁡(p′).q_0(p p)=q_0(p )=b(p ). (20) The base mass of cell j, accounting for validity, is therefore Bj=∑p∈Ωb(p)val(p)c(p)=j,ℛ=j:Bj>0,B_j= _p∈ b(p)val(p) 1\c(p)=j\, =\j:B_j>0\, (21) where ℛR is defined as the set of reachable cells. For any j∈ℛj , the within-cell target distribution is defined as πj(p)=b(p)val(p)c(p)=jBj. _j(p)= b(p)val(p) 1\c(p)=j\B_j. (22) For an active-cell index set ℐ⊆ℛI , the target distribution of a single emitted persona is the equal-cell mixture Πℐ=1|ℐ|∑j∈ℐπj,Πℐ(Ωj)=1|ℐ|(j∈ℐ). _I= 1|I| _j _j, _I( _j)= 1|I| (j ). (23) In contrast, the joint target distribution for the full population containing exactly one chain per cell is the product measure Γℐ=⨂j∈ℐπj. _I= _j _j. (24) Remark 5 (Distributional Distinctions). The distinction between πj _j, Πℐ _I, and Γℐ _I is critical: πj _j is a localized within-cell persona law, Πℐ _I is the desired marginal law of a single randomly emitted persona, and Γℐ _I is the stationary joint law of the entire population of parallel cell chains. The unknown masses BjB_j never need to be estimated, as they cancel from all within-cell Metropolis–Hastings ratios, while the parallel cell chains and the uniform scheduler inherently supply the equal mixture weights required by equation 23. Resolution-Limited Directional Uniformity. The following lemma formalizes the finite-resolution connection to true directional uniformity across the continuous semantic sphere. Lemma 4 (Wasserstein Bound on Uniformity). Let σℐ _I be the normalized spherical surface measure on ℐ=⋃j∈ℐjC_I= _j C_j, let νℐ=(W)#Πℐ _I=(z_W)_\# _I be the pushforward of Πℐ _I to the sphere, and let Δj _j be the geodesic diameter of cell jC_j. For every Wasserstein order qW≥1q_W≥ 1 WqW(νℐ,σℐ)≤(1|ℐ|∑j∈ℐΔjqW)1/qW≤maxj∈ℐ⁡Δj.W_q_W( _I, _I)≤ ( 1|I| _j _j^q_W )^1/q_W≤ _j _j. (25) Proof. Since both measures assign a probability mass 1/|ℐ|1/|I| to every individual cell jC_j, we can couple their conditional distributions separately within each respective cell. The physical geodesic distance between any two points coupled inside jC_j is naturally bounded by its diameter Δj _j. ∎ Lemma 4 connects equal cell masses to directional uniformity when cell diameters are small. Our implementation uses ten orthonormal sign cuts in 128 dimensions, to the shown bound does not provide a nontrivial approximation guarantee to uniform surface measure. Rather, the operational guarantee is equal allocation across reachable cells. C.1.2 Within-Cell Metropolis Kernels and Proposal Requirements Component Kernels and Acceptance Probabilities In addition to the global independence proposal q0q_0, let q1,…,qLqq_1,…,q_L_q be a set of fixed, exactly evaluable local-edit or block proposals. At each step, the algorithm first draws a proposal type ℓ with a fixed probability ωℓ>0 _ >0, such that ∑ℓ=0Lqωℓ=1 _ =0^L_q _ =1, and then applies a component-specific Metropolis–Hastings (MH) correction. For a chain assigned to an active cell j, the acceptance probability for transitioning from an incumbent persona p to a proposed persona p′p is αj,ℓ(p,p′)=p′∈Ωjmin1,b⁡(p′)qℓ(p∣p′)b⁡(p)qℓ(p′∣p). _j, (p,p )= 1\p ∈ _j\ \1,\, b(p )q_ (p p )b(p)q_ (p p) \. (26) A zero reverse probability implies a zero acceptance probability. Importantly, if the proposal type or edited block is selected with a state-dependent probability, that probability must be explicitly included in the forward–reverse ratio. Equation equation 26 describes a random mixture of separately corrected kernels; it does not represent the acceptance ratio for a single marginalized mixture proposal. Let Kj,ℓK_j, denote the resulting accept–reject kernel for proposal ℓ , which includes the self-transition upon rejection, and define the full mixture kernel as Kj=∑ℓ=0LqωℓKj,ℓ.K_j= _ =0^L_q _ K_j, . (27) Theorem 6 (Cell-Kernel Correctness and Matched Refresh). For every reachable cell j, each component kernel Kj,ℓK_j, is reversible with respect to πj _j; hence the mixture KjK_j leaves πj _j invariant. Moreover, under the matched global proposal (q0=bq_0=b), for every p∈Ωjp∈ _j, proposal value p′∈Ω⊥p ∈ _ , and measurable subset E⊆ΩjE _j αj,0(p,p′) _j,0(p,p ) =p′∈Ωj, = 1\p ∈ _j\, (28) Kj,0(p,E) K_j,0(p,E) =Bjπj(E)+(1−Bj)δp(E), =B_j _j(E)+(1-B_j) _p(E), (29) Kj(p,E) K_j(p,E) ≥ω0Bjπj(E). ≥ _0B_j _j(E). (30) where δp _p denotes the Dirac probability measure at p. Consequently, for every arbitrary initialization law ζ supported on Ωj _j and every integer n≥0n≥ 0 ‖ζKjn−πj‖TV≤(1−ω0Bj)n. \|ζ K_j^n- _j \|_TV≤(1- _0B_j)^n. (31) Proof. For any two valid same-cell states p,p′∈Ωjp,p ∈ _j, the two directional probability flows between them under kernel ℓ equal min⁡πj(p)qℓ(p′∣p),πj(p′)qℓ(p∣p′). \ _j(p)q_ (p p), _j(p )q_ (p p )\. Since this flow is perfectly symmetric, detailed balance is satisfied by the standard Metropolis–Hastings argument (Hastings, 1970; Tierney, 1994). When evaluating the global proposal (q0=bq_0=b), the proposal probabilities and base-weight terms cancel out for any valid same-cell move. A global proposal therefore hits Ωj _j with total probability BjB_j and, conditional on successfully doing so, is distributed according to law πj _j, yielding the independent refresh kernel Kj,0K_j,0 (equation 29). Because Kj,0K_j,0 is drawn with probability ω0 _0, the fixed mixture satisfies the minorization condition (equation 30). Finally, the resulting total-variation convergence bound follows from the classical Doeblin coupling argument (Mengersen & Tweedie, 1996; Rosenthal, 1995). ∎ Remark 6 (Global Cancellation and Strict Density Requirements). The cancellation that occurs for global proposals is mathematically beneficial, not degenerate: a same-cell hit results in an exact, independent refresh from the target πj _j. (The overall global acceptance rate may still appear small, because invalid and foreign-cell proposals remain as source-chain self-transitions.) However, for this exactness to hold, the forward and reverse proposal probabilities must describe the actual sampling law, including the EOS token and any block-selection probabilities. Therefore, the primary LLM implementation must sample raw categorical probabilities without top-kdeck_dec, top-pdecp_dec, retry-until-valid, or unaccounted many-to-one text normalization. If perfectly reversible local-edit blocks cannot be specified for free-form prose, the local component should be implemented exclusively as full-persona global MH proposal, ensuring the population-level random scan in equation 32 remains valid. C.1.3 The Hindsight-Spawning Population Algorithm. Let 0P_0 be the baseline persona population. We initialize the active-cell index set as the set of valid cells discovered in the baseline ℐ0=c(p):p∈0,val(p)=1.I_0=\c(p):p _0,\ val(p)=1\. For each j∈ℐ0j _0, we choose one arbitrary valid seed P(j)∈ΩjP^(j)∈ _j, assuming its base weight b⁡(P(j))>0b(P^(j))>0. This seed may instead be drawn uniformly from the baseline personas residing in that cell; its specific initialization law does not affect the anytime cell-level guarantees established below. Algorithm 3 Uniform-Coverage MCMC Persona Generation 1: Baseline population 0P_0; frozen evaluation components (val,W,c,b)(val,z_W,c,b); proposals qℓ=0Lq\q_ \_ =0^L_q and their probabilities ωℓ=0Lq\ _ \_ =0^L_q; total iterations TMT_M 2: Initialize ℐ0I_0 and one starting state P(j)∈ΩjP^(j)∈ _j for every j∈ℐ0j _0 3: for t=1,…,TMt=1,…,T_M do 4: Draw a target cell Jt|ℱt−1∼Uniform⁡(ℐt−1)J_t _t-1 (I_t-1) 5: Draw a proposal type Ht∼Categorical⁡(ω0,…,ωLq)H_t ( _0,…, _L_q) 6: Set p←P(Jt)p← P^(J_t) and draw a candidate Pt′∼qHt(⋅∣p)P _t q_H_t(· p) 7: Set ℐt←ℐt−1I_t _t-1 and leave the source chain P(Jt)P^(J_t) unchanged by default 8: if val⁡(Pt′)=1val(P _t)=1 then 9: j′←c⁡(Pt′)j ← c(P _t) 10: if j′=Jtj =J_t then 11: Draw Ut∼Uniform⁡(0,1)U_t (0,1) 12: if Ut≤αJt,Ht(p,Pt′)U_t≤ _J_t,H_t(p,P _t) then 13: P(Jt)←Pt′P^(J_t)← P _t 14: end if 15: else if j′∉ℐt−1j _t-1 then 16: Spawn a new chain P(j′)←Pt′P^(j )← P _t and expand the active set ℐt←ℐt−1∪j′I_t _t-1∪\j \ ⊳ hindsight seed; not an accepted source move 17: end if 18: end if 19: Emit the post-transition source state Ptemit←P(Jt)P_t^emit← P^(J_t) 20: end for 21: MCMC trace (P1emit,…,PTMemit)(P_1^emit,…,P_T_M^emit), final active population, and a separately labeled discovery archive Hindsight Mechanics and Rejection. A valid foreign-cell proposal is automatically rejected by the source chain because its source-cell target probability is zero. If its destination cell was previously inactive, however, it serves as a valid initialization for a new chain. This seed is not emitted immediately, nor is it recorded as an accepted MCMC transition for the source chain. A global hindsight seed entering a new cell j′j is distributed exactly according to πj′ _j ; a seed produced by a local-edit proposal, however, may have an arbitrary initialization law ζj′ _j supported on Ωj′ _j . Crucially, a proposal entering an already active foreign cell must not be submitted directly to the destination cell’s MH test, because the proposal was generated conditional on the source-chain’s state rather than the destination-chain’s state. Such cross-cell discoveries are instead retained in the separately labeled discovery archive. Finally, every scheduled proposal attempt advances the selected chain by one local MCMC iteration, even upon rejection. After a rejection, the emitted sample is the repeated current state, not the rejected candidate. (Discarding repeats or running until a fixed number of acceptances would improperly produce an accepted-state jump chain, violating the target measure. Fixed-Active-Set MCMC Interpretation. For a temporarily fixed active-cell set ℐI and a joint population state =(pj)j∈ℐp=(p_j)_j , we can define the full random-scan kernel as ℐ(,d′)=1|ℐ|∑j∈ℐKj(pj,dpj′)δ−j(d−j′). K_I(p,dp )= 1|I| _j K_j(p_j,dp _j) _p_-j(dp _-j). (32) Because each localized KjK_j preserves πj _j, the global kernel ℐ K_I preserves the product target Γℐ _I in equation 24. This maps to a standard random-scan, component-wise MCMC construction (Johnson et al., 2013): selecting a cell acts as the Gibbs coordinate step, and KjK_j acts as the within-coordinate MH update. At stationarity, emitting the selected coordinate yields the exact marginal law Πℐ _I. Hindsight spawning increases the number of active chains. With frozen components, however, the augmented process consisting of the active set ℐtI_t and its chain states is time-homogeneous on the disjoint union of population spaces. Before complete discovery, it is not the fixed-dimension random-scan chain associated with a fixed active-set product target. The theoretical guarantees in the following subsection are specifically designed to hold for this expansive target distribution. C.1.4 Theoretical Guarantees Standing Assumptions. To guarantee valid Markovian dynamics, all representation, validity, embedding, whitening, partition, target, and proposal components must be selected during a baseline or pilot phase and entirely frozen before the reported run. If the growing archive is allowed to change a proposal prompt, alter a validity rule, or update a target weight, the process becomes adaptive MCMC and; under such conditions, the fixed-kernel results established below no longer apply without imposing additional conditions (Roberts & Rosenthal, 2007). Theorem 7 (Anytime Active-Cell Uniformity). Let ℱt−1F_t-1 contain the complete history, the active set, and the individual chain states immediately prior to iteration t. For every initialization, t≥1t≥ 1, and j∈[M]j∈[M], the spatial allocation of the emitted persona is strictly uniform over the active set Pr⁡c⁡(Ptemit)=j∣ℱt−1=j∈ℐt−1|ℐt−1|. \c(P_t^emit)=j _t-1\= 1\j _t-1\|I_t-1|. (33) Moreover, define the predictable step exposure ηj,t _j,t, the cumulative realized count Nj(TM)N_j(T_M), and the cumulative expected exposure Ej(TM)E_j(T_M) as follows ηj,t _j,t =j∈ℐt−1|ℐt−1|, = 1\j _t-1\|I_t-1|, Nj(TM) N_j(T_M) =∑t=1TMc(Ptemit)=j, = _t=1^T_M 1\c(P_t^emit)=j\, Ej(TM) E_j(T_M) =∑t=1TMηj,t. = _t=1^T_M _j,t. (34) Then Nj(TM)−Ej(TM)N_j(T_M)-E_j(T_M) is a martingale. For every fixed horizon TMT_M and δ∈(0,1)δ∈(0,1), the deviation is strictly bounded Prmax1≤j≤M|Nj(TM)−Ej(TM)|≥TM2log⁡2Mδ≤δ. \ _1≤ j≤ M|N_j(T_M)-E_j(T_M)|≥ T_M2 2Mδ \≤δ. (35) Proof. The scheduler draws JtJ_t uniformly from ℐt−1I_t-1. Since the source chain either accepts a same-cell state or defaults to a self-transition, the relation c⁡(Ptemit)=Jtc(P_t^emit)=J_t almost surely, proving equation 33. Consequently, [c(Ptemit)=j∣ℱt−1]=ηj,tE[ 1\c(P_t^emit)=j\ _t-1]= _j,t, which gives the martingale. Applying the bounded-increment Hoeffding–Azuma inequality (Azuma, 1967) to each cell and taking union-bound over the fixed M cells yields equation 35. ∎ Remark 7 (Dynamic Allocation and Burn-In). The predictable exposure Ej(TM)E_j(T_M) is the only mathematically appropriate benchmark while the active set continues to grow, as a newly discovered cell cannot receive allocations retroactively. Thus, while the accumulated trace may not be uniform over the final set ℐTMI_T_M, it remains concentrated around the exact dynamic allocation induced by the uniform scheduler. Variance-adaptive confidence sequences can strengthen equation 35 to simultaneous validity over all stopping times (Howard et al., 2021); notably, no reset is required when a new cell is discovered. Furthermore, Theorem 7 acts as a precise no-burn-in guarantee: cell marginal uniformity is exact from the first emitted iteration. Burn-in remains relevant only if early personas are intended to be interpreted as draws from the asymptotic within-cell law πj _j, with that specific discrepancy bounded by equation 31. Theorem 8 (Eventual Discovery and Full-Target Convergence). Let τj=inft:j∈ℐt _j= \t:j _t\ be the discovery time of a reachable cell j∉ℐ0j _0. Under the matched global proposal and assuming ω0>0 _0>0 Pr⁡(τj>TM) ( _j>T_M) ≤(1−ω0Bj)TM, ≤(1- _0B_j)^T_M, (36) Pr⁡(ℐTM≠ℛ) (I_T_M ) ≤∑j∈ℛ∖ℐ0(1−ω0Bj)TM, ≤ _j _0(1- _0B_j)^T_M, (37) |ℛ∖ℐTM| |R _T_M| ≤∑j∈ℛ∖ℐ0(1−ω0Bj)TM. ≤ _j _0(1- _0B_j)^T_M. (38) Since ℛR is finite, every reachable cell is discovered in finite time almost surely. Let τ=inft:ℐt=ℛτ= \t:I_t=R\ and Mℛ=|ℛ|M_R=|R|. Conditional on the population at time τ, for every integer s≥0s≥ 0 the subsequent fixed-active-set chain obeys ‖ℛs(,⋅)−Γℛ‖TV≤min⁡1,∑j∈ℛ(1−ω0BjMℛ)s. \| K_R^\,s(p,·)- _R \|_TV≤ \1,\, _j (1- _0B_jM_R )^s \. (39) In particular, the emitted-persona law converges to the equal-cell mixture Πℛ=1|ℛ|∑j∈ℛπj. _R= 1|R| _j _j. (40) Proof. At each global iteration, the global proposal component is selected with probability ω0 _0 and independently draws from the base law b. It therefore proposes a valid persona in cell j with probability ω0Bj _0B_j, independently of the selected source cell. Repeated failure gives equation 36. Then, applying a union bound and the linearity of expectation give equation 37 and equation 38. Note that local proposals can only discover a cell earlier. The right-hand side of equation 37 tends to zero for finite ℛR, proving almost-sure eventual discovery. Following time τ, coordinate j is actively selected and undergoes an exact πj _j refresh with probability ω0Bj/Mℛ _0B_j/M_R per population iteration. By coupling a chain initiated from p with a stationary chain using the same scan, proposal, and accept–reject randomness, coordinate j coalesces at its first exact refresh and remains coupled thereafter. Union-bounding the probability that any coordinate has not refreshed by time s directly proves equation 39. To obtain convergence at deterministic times, let d⁡(s)d(s) denote the right-hand side of equation 39. The emission kernel QemitQ_emit acting on the pre-transition population satisfies ΓℛQemit=Πℛ _RQ_emit= _R. For any deterministic integer 0≤r<t0≤ r<t, conditioning on whether discovery is complete by time r and applying total-variation contraction gives ‖ℒ⁡(Ptemit)−Πℛ‖TV≤Pr⁡(τ>r)+d⁡(t−r−1). \|L(P_t^emit)- _R \|_TV≤ (τ>r)+d(t-r-1). Taking r=⌊t/2⌋r= t/2 makes both terms vanish as t→∞t→∞, proving equation 40. ∎ Remark 8 (The Resolution Trade-off and Interpretive Scope). The active coverage |ℐt||I_t| is nondecreasing, so each hindsight discovery is a deterministic coverage gain even though it also expands the target against which future allocation is measured. Equations equation 36–equation 38 quantify this gain. The discovery bounds also highlight the resolution trade-off: refinements to reduce cell diameters can tighten the geometric bound in equation 25. Under nested refinement, each child cell has a base mass no greater than its parent, which can slow discovery and refresh. Without a lower bound on the reachable cell masses, no distribution-free finite-time discovery guarantee can be operationally strong. Finally, we stress that the complete sequential trace, including all repeated states, is the object governed by these MCMC guarantees. A deduplicated archive composed of baseline states, accepted same-cell states, and valid cross-cell proposals is useful for measuring structural novelty and contrasting against empirical evolutionary generation, but it must be formally analyzed and reported as a visited-state archive rather than as an unweighted sample drawn from Πℐ _I. C.2 Evolutionary TextGrad Persona Generation Evolutionary TextGrad implements frontier generation by expanding the baseline population toward valid, low-density regions of persona space. Unlike UC-MCMC’s distributional target, this is an empirical search procedure; the resulting expanded pool is passed to the same max–min selector used for fixed-pool selection. Objective and textual feedback. We formulate persona generation as evolutionary constrained optimization. At iteration t, let tP_t be the current persona population. For any persona p, whether an incumbent or a newly proposed candidate, the comparison set is t∖pP_t \p\; set subtraction has no effect when p∉tp _t. Write ¯(p)=ϕ⁡(p)/‖ϕ⁡(p)‖2 e(p)=φ(p)/\|φ(p)\|_2 for the unit-normalized persona embedding. Let mE∈cos,2,Mahm_E∈\cos,2,Mah\ be the persona dissimilarity fixed for an evolutionary run; the reported run uses mE=Mahm_E=Mah. Define Gapt⁡(p) _t(p) =minp~∈t∖p⁡dmEP(p,p~), = _ p _t \p\d_m_E^P(p, p), (41) ρ^t(p) ρ_t(p) =1|t∖p|∑p~∈t∖pexp⁡(βKDE¯(p)⊤¯(p~)), = 1|P_t \p\| _ p _t \p\ \! ( _KDE\, e(p) e( p) ), (42) Fitt⁡(p) _t(p) =λrelVal⁡(p)+λgapGapt⁡(p)−λdenlog⁡ρ^t(p). = _relVal(p)+ _gapGap_t(p)- _den ρ_t(p). (43) Here Val⁡(p)∈[0,1]Val(p)∈[0,1] measures satisfaction of the persona schema and task-independent validity constraints, Gapt⁡(p)Gap_t(p) is nearest-neighbor novelty under the same dmEPd_m_E^P used for downstream persona selection, and ρ^t(p) ρ_t(p) is an unnormalized von Mises-Fisher kernel-density score in persona embedding space. Parent-path comparisons use an identical reference set for the parent and offspring. Thus, extend the definitions as Gapt⁡(x,R)Gap_t(x;R) and ρ^t(x,R) ρ_t(x;R) for a nonempty reference set R, where R=t∖xR=P_t \x\ without explicit reference set. Because personas are discrete text rather than differentiable vectors, we use textual gradient only for natural-language feedback, following Yuksekgonul et al. (2025) gttext(p)=TextGrad⁡(p,Val⁡(p),Gapt⁡(p),ρ^t(p),xgrad),g_t^text(p)=TextGrad\! (p;Val(p),Gap_t(p), ρ_t(p),x_grad ), (44) where xgradx_grad is the fixed feedback instruction. The editor then proposes p′=TGD.step⁡(p,gttext(p)).p =TGD.step\! (p,g_t^text(p) ). (45) This notation avoids treating LLM feedback as a literal derivative with respect to a token sequence. Evolutionary algorithm. The loop grows the population by mutating incumbents under textual-gradient feedback and admitting only offspring that demonstrably increase spread. Two design choices matter. First, selection and acceptance operate on the two diversity axes separately rather than on the scalar equation 43: a scalar trade-off would let a large novelty gain purchase an equally large density loss, which is exactly the substitution that produces a few far-flung outliers around an unchanged core. We therefore report FittFit_t as a summary statistic, while the loop itself is weight-free. Second, every admission score is evaluated against the frozen population t−1P_t-1, so individual scores do not depend on the order in which candidates are evaluated. Parent selection. Each tournament draws mtour=5m_tour=5 incumbents uniformly without replacement and admits two winners: the entrant with the largest Gapt−1Gap_t-1 and, when distinct, the entrant with the smallest log⁡ρ^t−1 ρ_t-1. Tournaments repeat until the parent set reaches βpar|t−1| _par|P_t-1| with βpar=0.05 _par=0.05, so the batch grows with the population. Admitting one winner per axis applies equal pressure to isolation and to sparsity without committing to an exchange rate between them. Admission. Let p be a parent and p′p its offspring, and let mgap=cmsdGapt−1⁡(q)q∈t−1,mden=cmsdlog⁡ρ^t−1(q)q∈t−1,m_gap=c_msd\Gap_t-1(q)\_q _t-1, m_den=c_msd\ ρ_t-1(q)\_q _t-1, (46) with cm=0.10c_m=0.10, be per-axis margins recomputed each iteration. The offspring is admitted when it passes the deterministic and relevance gates and satisfies either (parent path) (parent path) Gapt−1⁡(p′;Rp)>Gapt−1⁡(p;Rp)+mgap _t-1(p ;R_p)>Gap_t-1(p;R_p)+m_gap (47) log⁡ρ^t−1(p′,Rp)<log⁡ρ^t−1(p,Rp)−mden, ρ_t-1(p ;R_p)< ρ_t-1(p;R_p)-m_den, (absolute path) (absolute path) Gapt−1⁡(p′)≥Qτ _t-1(p )≥ Q_τ (48) log⁡ρ^t−1(p′)≤Q100−τ′, ρ_t-1(p )≤ Q_100-τ , where QτQ_τ and Q100−τ′Q _100-τ are the τ-th and (100−τ)(100-τ)-th population percentiles of the two axes, τ=75τ=75. The parent path requires improvement on both axes by explicit margins, discouraging admission based on negligible score differences. The absolute path admits a candidate that fails to beat its parent but is nonetheless competitive with the population, which prevents an already-isolated parent from blocking progress. Scoring is deliberately asymmetric between the two: the parent path scores p′p against Rp=t−1∖pR_p=P_t-1 \p\ so that parent and offspring are compared leave-one-out on identical reference sets, whereas the absolute path scores against the full population so that a near-copy of an elite parent cannot inherit that parent’s isolation. Sibling deduplication. Because offspring are scored against a frozen population, their admission scores do no account for other offspring of the same iteration. When two admitted offspring are closer than mgapm_gap, we treat them as near-duplicates and retain the one with the larger full-population gap. This step provides a heuristic control on redundancy within the admitted batch. Algorithm 4 Evolutionary TextGrad Persona Generation 1: Baseline population 0P_0; iteration cap TET_E; population cap NmaxN_ ; batch fraction βpar _par; tournament size mtourm_tour; margin fraction cmc_m; percentile τ; frozen feedback instruction xgradx_grad 2: for t=1,…,TEt=1,…,T_E do 3: Compute leave-one-out Gapt−1⁡(q)Gap_t-1(q) and log⁡ρ^t−1(q) ρ_t-1(q) for every q∈t−1q _t-1 4: Set the margins equation 46 and the percentile thresholds Qτ,Q100−τ′Q_τ,Q _100-τ 5: ℬt←∅B_t← 6: while |ℬt|<βpar|t−1||B_t|< _par|P_t-1| do 7: Sample mtourm_tour incumbents uniformly without replacement 8: Add the largest-GapGap entrant and, if distinct, the smallest-log⁡ρ ρ entrant 9: end while 10: t←∅A_t← 11: for p∈ℬtp _t do 12: p′←TGD.step⁡(p,gt−1text(p))p .step(p,g_t-1^text(p)) ⊳ equation 44–equation 45 13: if p′p fails the English or length gate or Val⁡(p′)≠1Val(p )≠ 1 then 14: reject ⊳ gates precede embedding and judging 15: else if equation 47 or equation 48 holds then 16: t←t∪p′A_t _t∪\p \ 17: end if 18: end for 19: Remove from tA_t the smaller-gap member of every pair within mgapm_gap ⊳ sibling deduplication 20: t←t−1∪tP_t _t-1 _t 21: if |t|≥Nmax|P_t|≥ N_ then 22: break 23: end if 24: end for 25: Evolved pool TEP_T_E; the max–min selector of Section 3.2 is applied downstream to return disp,mE⋆(TE)S _disp,m_E(P_T_E) The population is grown by union and never shrinks, so coverage of the persona space is monotone in t. Selection by max–min dispersion is deliberately kept downstream of the loop rather than folded into it: holding the selection operator fixed and varying only the candidate pool is what isolates the effect of generation from the effect of selection. Appendix D Experiment Details The number of personas k=5k=5: The Vendi score using cosine Gram matrix is measured as 10.410.4 on the baseline personas. This translates to an effective number of distinct personas in the baseline pool is 1010, and any number k≤10k≤ 10 should be able to accommodate k semantically distinct personas. We chose k=5k=5 for convenience and the feasibility of human clustering and judgment with multiple variations and baselines. Performance tables report means over five response-generation seeds (42–46) and two-sided 95% Student’s t confidence intervals. For seed-level metric values x1,…,x5x_1,…,x_5, computed using the benchmark-specific aggregation, we report x¯±t0.975,4s5,s2=14∑r=15(xr−x¯)2, x± t_0.975,4 s 5, s^2= 14 _r=1^5(x_r- x)^2, where t0.975,4≈2.776t_0.975,4≈ 2.776. These intervals summarize variability across response-generation seeds for the evaluated persona sets. ICC intervals in the judge-validation analysis use bootstrap resampling instead. D.1 Common Baselines and Variants Persona control baselines. Gibberish acts as a length-matched pseudoword control to verify that performance gains stem from semantic content rather than mere prompt length. Prompted with Jabberwocky personas, LLMs generate responses containing pseudowords (“shield from nasnorn rain” or “press flowai glurt sheevu”). Such out-of-vocabulary tokens embed in low-density regions, inflating every distance-based metric equation 57 without any corresponding remoteness of the underlying idea. We therefore delete them from each use before computing embeddings. The pseudowords originates from gibberish input: the gibberish generation process replaces content words one-for-one in place from an existing persona, so aligning a persona with its gibberish counterpart recovers a list fo non-words. A response word is deleted if it matches one of them, ignoring case and inflectional endings. The responses are evaluated after deletion (“shield from rain” or “press”). Task-conditioned persona generation (Jin et al., 2025) (MPAQ) serves as a baseline that explicitly generates task-specific personas. Task-specific generation inherently incurs expensive, per-task inference costs; by contrast, our generic diversified personas incur only a one-time optimization cost and can be deployed universally across any task. Due to this overhead, task-conditioned personas are not useful unless they are better for creative responses than task-agnostic personas. Baseline persona pool. We utilize the PersonaMem-v2 dataset (Jiang et al., 2025; Ge et al., 2024) as our fixed candidate pool. To extract k=5k=5 personas from this massive population, we evaluate two standard baselines: Random (uniform sampling) to test arbitrary individualization, and Typical (the 55 densest-mode personas via Gaussian KDE) to test representative personas. Diversity-selected personas (ours). Selection extracts k=5k=5 personas from the baseline pool by maximizing min dispersion (Section 3.2) under the Mahalanobis metric. Contrasting this with Random and Typical isolates the effect of diverse selection over standard individualization. Generated personas (ours). To break the performance ceiling imposed by a finite pool, we activaly expand the persona population using two algorithms prior to applying the identical max–min dispersion selector. MCMC acts as a space-filling generator targeting an equal-cell mixture over reachable semantic regions (Section 3.3). Evolution actively push personas toward low-density extremes by applying evolutionary TextGrad (Section 3.4). D.2 Alternative Uses Test (AUT) The Alternative Uses Test (AUT) is a widely used divergent-thinking benchmark (Guilford, 1967). Given a common object (e.g., a fork, book, or wallet), respondents produce alternative uses. Its open-ended, object-conditioned form has become a standard testbed for LLM creativity (Stevenson et al., 2022; Góes et al., 2023; Rabeyah et al., 2025) and is well suited to testing whether persona diversification broadens ideational range. For object o and experimental condition h, let ~o,h U_o,h denote the raw multiset of nraw=25n_raw=25 generated uses and let o,hU_o,h denote the canonicalized, deduplicated, valid set used by downstream metrics. Its actual size no,h=|o,h|n_o,h=|U_o,h| can therefore be smaller than 2525. D.2.1 Baselines and Variants We compare a graded ladder of prompting and persona conditions, each adding one ingredient over the last so that its marginal effect is isolated. All conditions share the generation protocol below; they differ only in the prompt and in how (or whether) a persona is supplied. Standard persona baselines. Three persona-free prompts from the ladder of Góes et al. (2023) isolate what prompting alone achieves: Common-Use (their n prompt, eliciting conventional uses) as a lower anchor; Alternative-Use (their nc “creative” prompt); and Creativity-Enhanced (their bs expert prompt with an explicit definition of creative use). Alternative-Use is the baseline for persona injection: every persona condition below supplies a persona to that same nc prompt, so any gain over a standard, non-persona prompt is attributable to the persona, not to prompt optimization. Evolutionary personas (ours). AUT-Evolution produces the pool of personas by the same evolutionary generation algorithm (Algorithm 4), but with AUT-specific fitness function. Fix the response generator and the benchmark prompt. For persona p and object o, let o(p)U_o(p) denote the valid uses that p produces for o under the protocol of Section D.2, and let mRm_R be the fixed response dissimilarity. A persona is scored on four axes, each averaged over the evaluated objects, Utilo⁡(p) _o(p) =1|o(p)|∑u∈o(p)so(u), = 1|U_o(p)| _u _o(p)s_o(u), (49) Novo⁡(p) _o(p) =1|o(p)|∑u∈o(p)minv∈ocomm⁡dmRR(u,v), = 1|U_o(p)| _u _o(p) _v _o^commd_m_R^R(u,v), (50) Divo⁡(p) _o(p) =(|o(p)|2)−1∑u,v⊆o(p)dmRR(u,v), = |U_o(p)|2^-1 _\u,v\ _o(p)d_m_R^R(u,v), (51) Flexo⁡(p) _o(p) =VS⁡(o(p)), =VS(U_o(p)), (52) so that novelty is a directed Chamfer distance to the frozen common-use reference set, diversity is within-persona spread, and flexibility is the effective number of distinct use categories equation 53. The utility strength so(u)s_o(u) is obtained by judging each use pairwise against K=3K=3 strength-stratified anchors drawn from a frozen per-object ladder and fitting a one-dimensional Bradley–Terry model against the fixed anchor strengths; each ladder is sum-zero identified, so UtiloUtil_o reads as an average advantage over baseline utility and is not comparable across objects. Validity is a gate rather than an axis: an offspring is discarded unless every retained use passes the parse, length, language, and semantic-admissibility checks of Section D.2. Write A⁡(p)=(Util,Nov,Div,Flex)(p)∈ℝ4A(p)=(Util,Nov,Div,Flex)(p) ^4 for the object-averaged axis vector. Admission replaces equation 47–equation 48 by a majority rule on the four axes: the offspring is accepted when it strictly beats its parent on at least w=3w=3 axes, or matches the τ-th percentile member of the population on at least w axes. Parent selection uses the same tournament, with the deciding axis rotated through a random permutation of the four so that each receives equal pressure over an iteration. Because all four axes are frozen within a metric epoch, the common-use pool, the per-object covariances, and the anchor ladder are snapshotted at run start, a later re-evaluation of the benchmark cannot retroactively move the objective a run was optimized against. Two properties of this substitution are worth stating. It is strictly more expensive, since each candidate must generate and have judged a full set of responses before it can be scored, whereas the persona-text fitness is computed from embeddings alone. And it is task-bound: the resulting pool is optimized for the AUT and carries no guarantee of transfer, which is precisely the trade the task-agnostic fitness avoids. We report both pools and compare them directly. Human reference. For the three objects with released human norms (book, fork, tin can), we include the Stevenson et al. (2022) human responses as a reference for the judge-scored dimensions. We sample ncomp=5n_comp=5 human respondents, matching the number of independent LLM completions per condition. Generation and sampling protocol. All responses are generated by Gemma-4-31B (temperature 0.70.7) under a uniform output constraint, five words per use, no adjectives, and an unbounded list, so conditions differ only in prompt and persona. The baseline population is 0=pii=1N0P_0=\p_i\_i=1^N_0 from PersonaMem (Jiang et al., 2025), with N0=997N_0=997 after excluding malformed or missing entries. If xox_o is the prompt instantiated for object o, a persona-conditioned response follows y∼GθR(⋅∣xo,p)y G_ _R(\,· x_o,p) with p∈Ω∅:=Ω∪∅p∈ _ := ∪\ \; persona-free conditions use the conditioning sentinel p=∅p= , which is distinct from the failed persona-generation symbol ⊥ . Non-persona conditions use ncomp=5n_comp=5 independent completions per object, while persona conditions use one completion for each of the k=5k=5 selected personas. We retain the first nuse=5n_use=5 uses from each completion, giving nraw=ncompnuse=25n_raw=n_compn_use=25 raw uses per object–condition pair before the validity gate. Variant Persona Description Common-Use – Góes et al. (2023) common-use (n) Alternative-Use – Góes et al. (2023) creative (nc) Creativity-Enhanced – Góes et al. (2023) expert (bs) Random ✓ 55 chosen uniformly at random Typical ✓ 55 density-ranked modes Selection (ours) ✓ maximize dispersion on base pool Evolution (ours) ✓ maximize dispersion on evolved pool Human – 55 chosen from Stevenson et al. (2022) Table 4: Variants, from persona-free baselines to our diversity-aware selection and generation. Persona variants all inject into the Alternative-Use prompt; Selection and Evolution share one selection operator and differ only in the candidate pool. D.2.2 Metrics Notation. Let ⊆Y denote the response-text space. The fixed encoder ϕ:→ℝdφ:T ^d maps any persona, prompt, object label, or response text to an embedding; write ⁡(⋅)=ϕ⁡(⋅)e(·)=φ(·). The regularized covariance Σ^R _R is estimated once on a frozen reference split and is not refit by condition. We call dcosRd_cos^R a dissimilarity because it need not satisfy the triangle inequality. Let ocommU_o^comm be the frozen common-use reference set for object o, drawn from no-persona baseline generations (Góes et al., 2023). For a valid response set =uii=1nU=\u_i\_i=1^n, let ¯i=⁡(ui)/‖⁡(ui)‖2 e_i=e(u_i)/\|e(u_i)\|_2, let E_U stack these unit vectors by row, and let =⊤G_U=E_UE_U be their cosine Gram matrix. If ξ1,…,ξn _1,…, _n are the eigenvalues of /nG_U/n, the Vendi Score is VS()=exp(−∑i=1nξilogξi),VS(U)= \! (- _i=1^n _i _i ), (53) with 0log⁡0=00 0=0. It is the effective number of distinct responses (Friedman & Dieng, 2023; Pasarkar & Dieng, 2024). Validity. We map each raw use through a deterministic canonicalizer can⁡(⋅)can(·) (lowercasing, filler and punctuation removal, and lemmatization). A separate, human-validated LLM judge supplies the admissibility indicator valR:→0,1val_R:Y→\0,1\. For ~=u~ii=1nraw U=\ u_i\_i=1^n_raw, Validity⁡(~)=1nraw∑i=1nrawvalR⁡(can⁡(u~i)).Validity( U)= 1n_raw _i=1^n_rawval_R\! (can( u_i) ). (54) The downstream set is =can(u~):u~∈~,valR(can(u~))=1,U= \can( u): u∈ U,\ val_R(can( u))=1 \, where set semantics merge exact canonical duplicates. Thus validity uses the raw denominator, whereas all following metrics use the post-gate denominator n=||n=|U|. Gating prevents invalid outliers from receiving spuriously high originality scores (Chen & Ding, 2023). Diversity. The first diversity metric is the mean pairwise response dissimilarity (defined for n≥2n≥ 2) Divm⁡()=2n⁡(n−1)∑i<jdmR(ui,uj).Div_m(U)= 2n(n-1) _i<jd_m^R(u_i,u_j). (55) Another diversity metric is the volume of convex hull on reduced PCA dimensions. Let em(u)∈ℝDme_m(u) ^D_m denote the response embedding realized in space m: the L2-normalized embedding for m=cosm=cos, the raw embedding for m=2m=2, and the whitened truncated embedding Lo⊤e~(u)L_o e(u) for m=Mahm=Mah, where Σo−1=LoLo⊤ _o^-1=L_oL_o is the Cholesky factorization of the object-specific inverse covariance. Let Πm,o:ℝDm→ℝk _m,o ^D_m ^k (k=5k=5) be the PCA projection fit on the pooled embeddings em(u)e_m(u) of all variants’ responses for object o. Then the volume is defined for n≥k+1n≥ k+1 as Volm⁡()=volk⁡(conv⁡Πm,oem(u1),…,Πm,oem(un)),Vol_m(U)=vol_k\! (conv \ _m,o\,e_m(u_1),…, _m,o\,e_m(u_n) \ ), (56) where conv⁡(⋅)conv(·) is the convex hull and volkvol_k the k-dimensional Lebesgue volume. Originality. We reserve originality for reference-relative response distance, distinguishing it from the persona novelty gap in equation 41. The object-relative and common-use-relative versions are Origm⁡(,o) _m(U;o) =1n∑i=1ndmR(ui,o), = 1n _i=1^nd_m^R(u_i,o), (57) Origm⁡(;ocomm) _m(U;U_o^comm) =1n∑i=1nminv∈ocomm⁡dmR(ui,v). = 1n _i=1^n _v _o^commd_m^R(u_i,v). The second expression is the directed Chamfer distance from the generated set to the frozen common-use reference set (Chen & Ding, 2023). Category Metrics (Flexibility and Surprise). Let γ⁡(u)γ(u) be the unique category assigned to valid use u by human clustering, based on action class (Fillmore, 1968; Levin, 1993) and affordance (Gibson, 1979; Norman, 1988), and define Cat⁡()=γ⁡(u):u∈Cat(U)=\γ(u):u \. The size-normalized manual-category metrics are Flexγ⁡()=|Cat⁡()|n,Flex_γ(U)= |Cat(U)|n, (58) Surγ(;ocomm)=1n∑i=1n[γ(ui)∉Cat(ocomm)].Sur_γ(U;U_o^comm)= 1n _i=1^n 1\! [γ(u_i) (U_o^comm) ]. (59) The clustering-free flexibility proxy is FlexVS⁡()=VS⁡()/nFlex_VS(U)=VS(U)/n. For the proposed originality-weighted Vendi proxy, set υi=minv∈ocomm⁡dcosR(ui,v),υ¯i=υi∑r=1nυr,¯=(υ¯1,…,υ¯n)⊤, _i= _v _o^commd_cos^R(u_i,v), _i= _i _r=1^n _r, =( _1,…, _n) , when ∑rυr>0 _r _r>0, and define (υ)=diag⁡(¯)diag⁡(¯).G_U^( )=diag( )\,G_U\,diag( ). If ξi(υ) _i^( ) are its eigenvalues, then SurVS(;ocomm)=1nexp(−∑iξi(υ)logξi(υ)).Sur_VS(U;U_o^comm)= 1n \! (- _i _i^( ) _i^( ) ). (60) We set this proxy to 00 when all υi=0 _i=0. Unlike SurγSur_γ, it is a proposed composite of reference-relative originality and within-set diversity, not a literal category share; its weighted spectral construction follows the quality-weighted Vendi framework (Nguyen & Dieng, 2024). Creativity Scores by LLM Judge. The creativity, originality, surprise, and utility of each use are rated by an LLM judge. Following Góes et al. (2023), all valid canonical uses for object o are pooled across conditions, randomly partitioned into batches, ranked and scored, and repeated for Rjudge=5R_judge=5 rounds with fresh partitions. The rank score is the mean within-batch percentile, and the judge score is the mean raw rating, scaled to [1,5][1,5]. With the same judge, we obtain absolute per-use ratings of originality, surprise, and utility using faithful renderings of Stevenson et al. (2022)’s human protocols. Utility is reported separately but remains the effectiveness component of creativity, rather than an unrelated quantity (Runco & Jaeger, 2012). Metric Basis Description Validity Human / LLM judge Proportion of admissible, well-formed uses (gate; control) Diversity Embedding Mean pairwise distance among a variant’s uses (↑ ) Originality Embedding Mean semantic or directed Chamfer distance from uses to the object concept or common uses (↑ ) Flexibility Category (Effective) number (ratio) of distinct use categories (↑ ) Surprise Category (Effective) share of uses in categories beyond common uses (↑ ) Creativity LLM judge Pooled rank-and-score creativity, Góes et al. (2023) (↑ ) Originality LLM judge Deviation from typical use, Stevenson et al. (2022) (↑ ) Surprise LLM judge Unexpectedness, Stevenson et al. (2022) (↑ ) Utility LLM judge Feasibility/usefulness (control), Stevenson et al. (2022) Table 5: Creativity and diversity metrics. Embedding metrics use the fixed encoder ϕφ (EmbeddingGemma) with cosine, L2, or Mahalanobis dissimilarity; category metrics use both human clustering and Vendi-based effective counts; LLM-judge metrics use an independent Qwen3.6-27B judge. ↑ : higher indicates a larger value. Validity is a gate. Utility is the effectiveness dimension required alongside originality for creativity (Runco & Jaeger, 2012) and is reported separately to expose trade-offs. We use Qwen3.6-27B, an open-weight model from a family distinct from the Gemma-4-31B generator. Qwen3.6-27B shows moderate agreement with the two-rater mean on Stevenson-style ratings (Spearman (ρ=.566ρ=.566)–(.635.635); ICC((C,1)=.407(C,1)=.407)–(.568.568)), below the human–human agreement reference ((ρ=.669ρ=.669)–(.842.842); ICC (=.677=.677)–(.790.790)). We therefore interpret these scores as noisy ordinal indicators, with strongest support for originality, and rely on convergence with embedding- and category-based measures rather than treating them as interchangeable with human ratings. Dim. Pair n ICC(C,1)(C,1) 95% CI QWK ρ Exact Within-1 Paper ICC Orig. LLM vs. rater 1 139 0.480 [0.341,0.609][0.341,0.609] 0.447 0.553 0.547 0.878 0.78 Orig. LLM vs. rater 2 139 0.575 [0.452,0.693][0.452,0.693] 0.574 0.599 0.504 0.942 0.78 Orig. LLM vs. mean 139 0.568 [0.443,0.682][0.443,0.682] 0.481 0.635 0.302 0.878 0.78 Orig. Human ceiling 139 0.677 [0.577,0.754][0.577,0.754] 0.609 0.669 0.475 0.993 0.78 Util. LLM vs. rater 1 139 0.355 [0.211,0.504][0.211,0.504] 0.344 0.509 0.475 0.813 0.70 Util. LLM vs. rater 2 139 0.430 [0.292,0.568][0.292,0.568] 0.406 0.597 0.532 0.806 0.70 Util. LLM vs. mean 139 0.407 [0.281,0.531][0.281,0.531] 0.377 0.603 0.396 0.784 0.70 Util. Human ceiling 139 0.701 [0.493,0.851][0.493,0.851] 0.696 0.745 0.712 0.978 0.70 Surp. LLM vs. rater 1 139 0.432 [0.229,0.608][0.229,0.608] 0.423 0.500 0.612 0.935 0.79 Surp. LLM vs. rater 2 139 0.544 [0.407,0.659][0.407,0.659] 0.490 0.571 0.525 0.885 0.79 Surp. LLM vs. mean 139 0.523 [0.354,0.667][0.354,0.667] 0.468 0.566 0.446 0.906 0.79 Surp. Human ceiling 139 0.790 [0.728,0.841][0.728,0.841] 0.756 0.842 0.698 0.971 0.79 Table 6: AUT LLM-as-a-judge agreement on the valid Stevenson subset (n=139n=139). Paper ICC is the human–human ICC reported by Stevenson et al.; Human ceiling is the reproduced rater01–rater02 agreement on the released file. LLM scores are from Qwen3.6-27B with the faithful Stevenson prompts (including 0/99 codes). Dim. Pair n ICC(C,1)(C,1) 95% CI QWK ρ Exact Within-1 Paper ICC Orig. LLM vs. rater 1 147 0.386 [0.248,0.525][0.248,0.525] 0.378 0.488 0.524 0.844 0.57 Orig. LLM vs. rater 2 147 0.479 [0.341,0.608][0.341,0.608] 0.474 0.533 0.476 0.905 0.57 Orig. LLM vs. mean 147 0.467 [0.332,0.601][0.332,0.601] 0.410 0.555 0.286 0.837 0.57 Orig. Human ceiling 147 0.630 [0.498,0.734][0.498,0.734] 0.579 0.642 0.476 0.986 0.57 Util. LLM vs. rater 1 147 0.336 [0.199,0.472][0.199,0.472] 0.316 0.499 0.449 0.769 0.68 Util. LLM vs. rater 2 147 0.400 [0.264,0.534][0.264,0.534] 0.367 0.572 0.503 0.769 0.68 Util. LLM vs. mean 147 0.382 [0.262,0.501][0.262,0.501] 0.362 0.584 0.374 0.748 0.68 Util. Human ceiling 147 0.684 [0.505,0.830][0.505,0.830] 0.680 0.750 0.694 0.973 0.68 Surp. LLM vs. rater 1 147 0.391 [0.210,0.566][0.210,0.566] 0.374 0.455 0.585 0.905 0.67 Surp. LLM vs. rater 2 147 0.480 [0.344,0.602][0.344,0.602] 0.423 0.501 0.497 0.850 0.67 Surp. LLM vs. mean 147 0.467 [0.313,0.613][0.313,0.613] 0.411 0.502 0.422 0.878 0.67 Surp. Human ceiling 147 0.772 [0.704,0.832][0.704,0.832] 0.743 0.831 0.694 0.966 0.67 Table 7: AUT LLM-as-a-judge agreement on all doubly rated Stevenson responses (n=147n=147), including rows with invalid codes. Metrics as in Table 6. Interpretation. On the production-relevant valid subset, the LLM tracks the human mean with Spearman ρ∈[0.57,0.64]ρ∈[0.57,0.64] and within-one rates of 7878–91%91\%: most disagreements are off-by-one on a five-point scale rather than ordinal inversions. Relative to the reproduced human–human ceiling, the LLM recovers about 84%84\% of the ceiling ICC and 95%95\% of the ceiling Spearman on originality, and still 5858–81%81\% of the ceiling on utility and surprise. Exact agreement with the continuous rater mean is lower by construction (integer LLM scores vs. half-point means), so we emphasize ICC, QWK, Spearman, and within-one. Taken together, the judge is below, but in the same regime as trained human reliability, which is the appropriate bar for LLM-as-a-judge deployment. We therefore treat the Stevenson originality, utility, and surprise scores in the main AUT results as credible automated ratings, with originality the best-aligned dimension and utility the most conservative. D.3 Infinity-Chat 100 We evaluate five manually selected queries from Infinity-Chat 100 Jiang et al. (2026): 1, 2, 46, 60, and 78. For each query and condition, persona-free prompting produces 50 responses, while persona conditions produce ten responses from each of five personas. This yields 250 responses per condition per generation seed. Generation uses temperature 1.01.0, top-p=0.9p=0.9, and a maximum of 512 new tokens. Before computing embedding and lexical metrics, we exclude responses explicitly labeled invalid. D.3.1 Baselines and Variants The persona-free baseline uses the original query. We also evaluate the shared ZS-CoT, Step-Back, and DMAD prompt variants, together with the persona controls and selection and generation variants defined above. The composition experiment adds the selected personas to the DMAD prompt. Task-specific evolutionary baseline. IC-Evolution replaces persona-space fitness with three response-space axes: mean pairwise cosine distance, Vendi score, and one minus mean word-trigram Jaccard overlap. These are computed within each query and averaged across queries after validity screening and mechanical removal of persona framing. The resulting population is supplied to the downstream Mahalanobis max–min dispersion selector. D.3.2 Metrics Homogeneity. For a query and condition, let =(u1,…,un)U=(u_1,…,u_n) denote the response collection, and write dij=dcosR(ui,uj)d_ij=d_cos^R(u_i,u_j). Homogeneity is mean pairwise cosine similarity, Hom⁡()=1−2n⁡(n−1)∑i<jdij.Hom(U)=1- 2n(n-1) _i<jd_ij. Persona Separation. Let pip_i be the persona generating uiu_i. Define ℬ=(i,j):i<j,pi≠pjB=\(i,j):i<j,\ p_i≠ p_j\ and ℐ=(i,j):i<j,pi=pjI=\(i,j):i<j,\ p_i=p_j\. Persona separation is Sep⁡()=1|ℬ|∑(i,j)∈ℬdij−1|ℐ|∑(i,j)∈ℐdij.Sep(U)= 1|B| _(i,j) d_ij- 1|I| _(i,j) d_ij. Separation is undefined when either pair set is empty. Both metrics are computed within each query and averaged across queries. Flexibility. We apply the cosine-Gram Vendi score in equation 53 to the response embeddings for each query and average across queries. Validity and Quality. Validity is the fraction of responses admitted by the validity screen. Qwen3.6-27B independently rates the quality of each response on a five-point scale for execution and fit to the request and the scores are averaged across responses and queries. LLM Judge on Response Quality and Validity Infinity-Chat releases human ratings of model responses without human-written response pools Jiang et al. (2026). We use the ratings to validate Qwen3.6-27B judge employed in evaluation: 750750 (query, response) pairs from 2525 external LMs, each with approximately 2525 independent absolute ratings on a 1–5 scale (human_absolute). The quality head is scored against the 2525-rater mean (the paper’s comparison target for LM judges); the validity head is checked with a tripwire that responses humans rate ≥4≥ 4 must almost never be labeled invalid. Table 8 reports rank and absolute agreement for quality; Table 9 reports the validity screen. Metric Value n scored pairs 750 Quality prompt hivemind_quality_absolute_v1 Spearman ρ (judge vs. human mean) 0.158 Pearson r (judge vs. human mean) 0.278 ICC(C,1)(C,1) 0.225 ICC 95% CI (bootstrap) [0.146, 0.296][0.146,\,0.296] MAE 1.021 RMSE 1.156 Mean bias (judge −- human mean) +0.691+0.691 Exact agree (vs. rounded human mean) 0.136 Within-one (vs. rounded human mean) 0.840 QWK (vs. rounded human mean) 0.132 Human split-half Spearman ceiling 0.624 ρ / ceiling 0.253 Table 8: Infinity-Chat quality-judge agreement against the 2525-rater human mean on the Artificial Hivemind dense absolute-rating set. Exact / within-one / QWK compare the integer LLM score to the human mean rounded to the nearest integer. Human mean bin n Invalid rate Prompt Good (≥4≥ 4) 405 0.002 hivemind_validity_screen_v1 Mid ((2,4)(2,4)) 341 0.023 Bad (≤2≤ 2) 4 0.250 Overall 750 0.013 Table 9: Infinity-Chat validity-gate tripwire on the same 750750 pairs. The automated gate fails the protocol if it marks >5%>5\% of human-good responses invalid; the observed rate is 0.25%0.25\%. The validity screen is highly conservative with respect to human judgments: only 0.25%0.25\% of responses that humans rate as good (≥4≥ 4) are discarded, well under the 5%5\% tripwire, so the LLM gate does not systematically delete content that human raters endorse. On quality, the judge is positively associated with the human consensus (Pearson r=0.28r=0.28, ICC=0.23=0.23) and lands within one scale point of the rounded human mean on 84%84\% of items, while remaining below the human split-half Spearman ceiling (ρ=0.16ρ=0.16 vs. 0.620.62). The residual gap is consistent with a mild high-score bias (mean bias +0.69+0.69) on a distribution where human means already cluster near the top of the scale. We therefore treat the Infinity-Chat validity labels as a credible hard filter for all downstream LLM and embedding metrics, and treat quality scores as a coarsely human-aligned complementary axis, which is useful for relative quality, diversity tradeoffs across variants, and interpreted alongside embedding-based homogeneity and diversity rather than as a calibrated substitute for the full 2525-rater panel. D.4 Divergent Association Task (DAT) D.4.1 Baselines and Variants Standard persona baselines. We use the original DAT prompt (Olson et al., 2021; Chen & Ding, 2023) as the baseline persona-free prompt: DAT. Schapiro et al. (2026) tests various prompts to contrast creative and non-creative outputs, and we also compare their task-optimized prompts as our baseline comparisons: CN-Base (their base instruction prompt that contrasts with random instruction prompt), CN-Random (their random instruction prompt), Creativity-enhanced (their creative prompt that serves as a contrast to their non-creative prompt), Non-creative (their non-creative prompt that serves the similar role as common use prompt of AUT benchmark). Other baselines and variants. Other reasoning baselines, persona controls (gibberish, Jin et al. (2025), and random/typical personas from Jiang et al. (2025); Ge et al. (2024)) are similarly compared for DAT as well. Our selection and generation (MCMC and Evo) variants use the same set of selected personas used for AUT. (Note that the personas, as well as selection and generation algorithms are agnostic of a specific task). DAT-Evolution is the task-conditioned variant using DAT-specific fitness function Fit(p)=λDATDATm(1:7)+λFlexVS()+λH()+λC(1−C10()),Fit(p)= _DATDAT_m(W_1:7)+ _FlexVS(W)+ _HH(W)+ _C(1-C_10(W)), (61) where W denotes the set of valid responses produced by the persona p. Generation and sampling protocol. Non-persona conditions use ncomp=35n_comp=35 independent completions, while persona conditions use seven completions for each of the k=5k=5 selected personas. We retain the first nword=10n_word=10 words from each completion, giving nraw=350n_raw=350 raw words per condition before validity gate. D.4.2 Metrics For DAT, we report four creativity metrics along with validity ratio: DAT score as suggested by the original literature (Olson et al., 2021), fluency measured by Shannon entropy, flexibility measured by Vendi score equation 53, and top-10 concentration. Notation. For DAT, we denote response as w and a valid response set as W, as they represent noun words. Let Wj=(wj1,…,wjnj)W_j=(w_j1,…,w_jn_j) contain the usable validated word occurrences from completion j in response order. Unlike AUT, the relationship between 1010 words in a completion is also an important evaluation target, because the task is to name 1010 irrelevant nouns. Thus, we deal with the pooled multiset =⨄jWjW= _jW_j. Let humanW^human be the human reference set of words, drawn from human responses collected by Olson et al. (2021). DAT Score. Olson et al. (2021) invented Divergent Association Task (DAT) along with a creativity metric called DAT score. It is adopted in its original form by following works using DAT (Chen & Ding, 2023; Schapiro et al., 2026). Let g¯(w) g(w) denote its unit-normalized GloVe word embedding. For a scorable completion (nj≥7n_j≥ 7), DAT⁡(Wj)=100(nj2)∑1≤a<b≤nj[1−g¯(wja)⊤g¯(wjb)].DAT(W_j)= 100 n_j2 _1≤ a<b≤ n_j [1- g(w_ja) g(w_jb) ]. Human-reference percentiles use the same calculation applied to the raw word lists of 8,572 human completions. Flency. Shannon entropy captures fluency on how flat the generated words are distributed: higher entropy, lower duplicate words and thus higher fluency. Therefore, the metric measures how creative/divergent LLMs are H()=−∑w∈cw∑w∈cwlogcw∑w∈cw,H(W)=- _w c_w _w c_w c_w _w c_w, (62) where V is vocabulary or the set of unique words in W and cwc_w is the count of appearance of w in W. Top-10 Concentration measures fluency on how much proportion the top-10 most frequent words appear over and over again: lower top-10 share, lower concentration and thus higher fluency. Therefore, the metric measures how lexically sophisticated LLMs are C10()=∑i=110c(i)∑w∈cw,C_10(W)= _i=1^10c_(i) _w c_w, (63) where c(i)c_(i) is the count of appearance i−i-th most frequently appearing word in W. Flexibility. For each nonempty completion WjW_j, we apply equation 53 to the unit-normalized EmbeddingGemma embeddings of its individual usable word occurrences, retaining repeated words as repeated embedding rows. The reported Vendi score is the mean of these within-completion scores over nonempty completions. It measures effective semantic diversity within a generated word list, whereas type token ratio (TTR) is the ratio of unique valid words to all valid word occurrences TTR()=||||.TTR(W)= |V||W|. (64) Appendix E Experiment Results E.1 Baselines E.1.1 Common Baselines Here we evaluate and compare various metrics on 12 categories of variants: 3 standard persona baselines using reasoning prompts, 2 persona control baselines, 2 PersonaMem-v2 (Jiang et al., 2025) baselines, 2 selection algorithm variants, and 3 generation algorithm variants. These are existing baselines using reasoning prompts: ▶ ZS-CoT: Zero-Shot Chain-of-Thought (ZS-CoT) reasoning prompt adapted from Kojima et al. (2022) ▶ Step-Back: Step-Back Prompting (SBP) adapted from Zheng et al. (2024) ▶ DMAD: Diverse Multi-Agent Debate prompt combining reasoning strategies (ZS-CoT + SBP) adapted from Liu et al. (2025) We also compare our methods against control personas: ▶ Gibberish: gibberish of equivalent length as personas our methods utilize ▶ Task-conditioned: task-conditioned personas generated by Jin et al. (2025) We compare against existing persona benchmark PersonaMem-v2 dataset (Jiang et al., 2025), which randomly selected from PersonaHub (Ge et al., 2024), using two sampling methods: ▶ Random: randomly selected personas from the base population. ▶ Typical: typical personas selected from the most dense regions. Finally, we apply our diversity-aware selection methods on persona population of Jiang et al. (2025): ▶ Coverage: selected personas from the base population to maximize coverage defined by cosine, L2 and Mahalanobis distances. ▶ Dispersion: selected personas from the base population to maximize minimum dispersion defined by cosine, L2 and Mahalanobis distances; Mahalnobis distance-based dispersion selection is chosen as a default for all persona generation algorithms. Then, expand the baseline population using our MCMC and evolutionary persona generation algorithms: ▶ MCMC: extended population using Uniform-Coverage MCMC persona generation algorithm. ▶ Evolution: extended population using Evolutionary TextGrad algorithm on persona fitness function. ▶ AUT Evolution: extended population using Evolutionary TextGrad algorithm on AUT response fitness function. E.1.2 AUT Baselines These are the baseline and task-optimized prompts from Góes et al. (2023): ▶ Common Use: common uses of objects ▶ Alternative Use: alternative uses of objects generated by generic AUT prompt ▶ Expert: alternative uses generated by AUT-optimized expert prompt ▶ Creativity-enhanced: alternative uses generated by creativity-enhanced prompt adapted from Góes et al. (2023) E.1.3 DAT Baselines These are the prompt conditions for the DAT, following Chen & Ding (2023) and Schapiro et al. (2026): ▶ Divergent Association: zero-shot task prompt in the original form of Olson et al. (2021), adapted by Chen & Ding (2023). ▶ Base-Instruction: base control prompt (Chen & Ding, 2023), which asks for ten nouns with no divergence instruction. ▶ Random-Instruction: random control prompt (Chen & Ding, 2023), which asks for ten random nouns. ▶ Creative / Non-Creative (Non-Divergent Association): the contrastive prompt pair of Schapiro et al. (2026), which instruct the model to answer creatively and uncreatively respectively, bracketing the achievable range. Every condition beyond DAT, the reasoning prompts, the control personas, and all of our selection and generation variants, injects into the Divergent Association prompt, which is therefore the reference condition throughout the DAT results. Selection under cosine and L2L_2 dissimilarity returns identical persona sets at every pool, so the two are reported jointly. E.2 Persona Diversity Table 10: Persona diversity of the variants measured in Dispersion (equation 8) and Vendi score (equation 53) of each persona set (k=5k=5), computed on persona embeddings under cosine, L2, and Mahalanobis dissimilarity. Dispersion represents how personas are spread far from each other (minimum pairwise distance within the set; higher is more spread). Vendi score represents the effective number of distinct clusters of personas (higher is more distinct). Methods are grouped by their algorithm categories: MPAQ is the baseline (Jin et al., 2025) that generated persona for output diversity, Individualized personas represent baseline selection methods from the base pool of size 997, Selected personas represent our methods of persona selection according to various objective functions, and Generated personas represent our methods of persona generation, each selected from its expanded pool by exact Mahalanobis max–min dispersion. Bold = best per column within the selected and generated groups; baseline groups are not bolded. Persona set Dispersion Vendi score Cosine L2 Mahal. Cosine L2 Mahal. Baseline personas Task-conditioned 0.1300.130 0.5100.510 13.4213.42 1.811.81 2.212.21 1.211.21 Individualized personas (base pool =997=997) Typical 0.2700.270 0.7340.734 12.0412.04 2.342.34 2.732.73 1.211.21 Random 0.2480.248 0.7030.703 13.6713.67 2.602.60 2.922.92 1.251.25 Selected personas (base pool =997=997) Coverage, cosine 0.2310.231 0.6790.679 12.2512.25 2.222.22 2.632.63 1.191.19 Coverage, Mahalanobis 0.1910.191 0.6190.619 10.5210.52 2.252.25 2.642.64 1.151.15 Dispersion, cosine 0.4850.485 0.9840.984 16.0516.05 3.413.41 3.413.41 1.381.38 Dispersion, Mahalanobis 0.3660.366 0.8560.856 20.9820.98 3.073.07 3.233.23 1.461.46 Generated personas (expanded pool size) MCMC (8,553) 0.3520.352 0.8390.839 20.9820.98 3.043.04 3.213.21 1.461.46 Evolution (1,869) 0.3930.393 0.8870.887 24.3824.38 3.193.19 3.303.30 1.641.64 AUT-Evolution (3,397) 0.3730.373 0.8640.864 22.3422.34 3.123.12 3.263.26 1.511.51 DAT-Evolution (2,439) 0.4360.436 0.9340.934 21.6121.61 3.283.28 3.343.34 1.491.49 Figure 5: Persona-to-response relationship on (a) the Alternative Uses Task and (b) Infinity-Chat. Points represent 14 analogous task-agnostic method configurations per benchmark; response hull extent and vendi score are averaged over seven AUT objects and five Infinity-Chat queries. Arrows connect six matched Coverage→ contrasts while holding candidate-pool family and selector metric fixed. Dashed lines show configuration-level linear fits; annotations report Pearson correlations and descriptive 95% intervals. E.3 AUT Benchmark Table 11: Response validity: the proportion of generated uses admitted by the validity gate of Section D.2, reported over all seven AUT objects and, for comparability with Stevenson, over the three objects with human reference data (175175 and 7575 generated uses per variant, respectively). Every variants beyond Alternative-Use build prompts upon the generic Alternative-Use prompt unless its group states otherwise; the final two groups re-run our selection and generation methods on the Creativity-enhanced prompt of Góes et al. (2023), showing that persona diversification composes with a stronger prompting strategy rather than competing with it. Cells give the admitted proportion averaged over objects. Subscripts denote the 95% confidence interval across 5 independent random seeds. Stevenson exists only for the human-reference objects, so its seven-object cell is omitted (—). Validity is a control rather than a target: values near 11 are expected and the informative signal is degradation rather than rank, so no value is bolded. Generated variants report their expanded pool size in parentheses. Variant All seven objects Human-reference objects Standard persona Common-Use 0.971±0.0180.971_± 0.018 0.981±0.0150.981_± 0.015 Alternative-Use 0.984±0.0140.984_± 0.014 0.968±0.0250.968_± 0.025 Expert 0.997±0.0060.997_± 0.006 0.997±0.0070.997_± 0.007 Creativity-enhanced 0.968±0.0110.968_± 0.011 0.981±0.0220.981_± 0.022 ZS-CoT 0.978±0.0090.978_± 0.009 0.997±0.0070.997_± 0.007 Step-Back 0.976±0.0140.976_± 0.014 0.968±0.0150.968_± 0.015 DMAD 0.977±0.0090.977_± 0.009 0.997±0.0070.997_± 0.007 Gibberish 0.693±0.0350.693_± 0.035 0.659±0.1250.659_± 0.125 Baseline personas MPAQ 0.930±0.0240.930_± 0.024 0.947±0.0200.947_± 0.020 Individualized personas Random 0.993±0.0060.993_± 0.006 0.989±0.0140.989_± 0.014 Typical 0.995±0.0060.995_± 0.006 0.997±0.0070.997_± 0.007 Selected personas from base pool Coverage, cosine 0.987±0.0060.987_± 0.006 1.000±0.0001.000_± 0.000 Coverage, Mahalanobis 0.993±0.0130.993_± 0.013 1.000±0.0001.000_± 0.000 Dispersion, cosine 0.992±0.0120.992_± 0.012 0.992±0.0150.992_± 0.015 Dispersion, Mahalanobis 0.986±0.0100.986_± 0.010 0.995±0.0090.995_± 0.009 Generated personas (expanded pool size) MCMC (8,553) 0.984±0.0090.984_± 0.009 0.995±0.0090.995_± 0.009 Evolution (1,869) 0.985±0.0120.985_± 0.012 0.987±0.0170.987_± 0.017 AUT-Evolution (3,397) 0.978±0.0170.978_± 0.017 0.989±0.0140.989_± 0.014 Selected personas ++ Creativity-enhanced prompt Coverage, cosine 0.978±0.0120.978_± 0.012 0.981±0.0150.981_± 0.015 Coverage, Mahalanobis 0.989±0.0140.989_± 0.014 0.992±0.0090.992_± 0.009 Dispersion, cosine 0.982±0.0080.982_± 0.008 0.979±0.0190.979_± 0.019 Dispersion, Mahalanobis 0.979±0.0110.979_± 0.011 0.976±0.0300.976_± 0.030 Generated personas ++ Creativity-enhanced prompt MCMC 0.967±0.0140.967_± 0.014 0.968±0.0400.968_± 0.040 Evolution 0.979±0.0200.979_± 0.020 0.973±0.0310.973_± 0.031 AUT-Evolution 0.973±0.0060.973_± 0.006 0.968±0.0340.968_± 0.034 Human Stevenson — 0.811±0.0070.811_± 0.007 Table 12: Response diversity across all seven AUT objects, reported two ways. Mean pairwise distance (equation 55) is the average cosine, L2 or Mahalanobis distance among the valid uses of a variant; convex hull volume (equation 55) is the volume of the convex hull those uses span in the same geometry, so it rewards a set that occupies a region rather than merely separating in the mean. Both are averaged over objects. Subscripts denote the 95% confidence interval across 5 independent random seeds, and higher is more diverse. Hull extent under cosine and L2 coincides to nine decimal places, the two metrics are monotone transforms of one another on unit-normalized embeddings, so it is reported once. Baseline and our method variants are grouped by the personas: standard persona represents LLMs default persona when no persona is given explicitly, baseline personas represent existing persona injection methods, individualized personas represent broad population of personas from PersonaHub Ge et al. (2024); Jiang et al. (2025), while selected personas and generated personas represent our specific persona selection and generation algorithms. The final two groups re-run our selection and generation methods on the Creativity-enhanced prompt, showing that persona diversification composes with a stronger prompting strategy. Bold marks the best value per column within a method category, with tied values bolded jointly. All personas, including gibberish, are injected to the alternative use prompt unless their group states otherwise. Mean pairwise distance Convex hull volume Variant Cosine L2 Mahal. Cosine/L2 Mahal. Standard persona Common-Use 0.155±0.0040.155_± 0.004 0.526±0.0060.526_± 0.006 15.50±0.3515.50_± 0.35 0.103±0.0040.103_± 0.004 2.17±0.232.17_± 0.23 Alternative-Use 0.188±0.0030.188_± 0.003 0.590±0.0060.590_± 0.006 11.97±0.2811.97_± 0.28 0.172±0.0140.172_± 0.014 1.42±0.091.42_± 0.09 Expert 0.179±0.0040.179_± 0.004 0.580±0.0060.580_± 0.006 14.57±0.2914.57_± 0.29 0.161±0.0040.161_± 0.004 2.36±0.132.36_± 0.13 Creativity-enhanced 0.179±0.0020.179_± 0.002 0.592±0.0040.592_± 0.004 20.24±0.5620.24_± 0.56 0.144±0.0100.144_± 0.010 5.15±0.165.15_± 0.16 ZS-CoT 0.165±0.0030.165_± 0.003 0.563±0.0060.563_± 0.006 14.73±0.3814.73_± 0.38 0.161±0.0060.161_± 0.006 2.43±0.112.43_± 0.11 Step-Back 0.173±0.0060.173_± 0.006 0.574±0.0140.574_± 0.014 13.66±0.5213.66_± 0.52 0.168±0.0110.168_± 0.011 2.03±0.292.03_± 0.29 DMAD 0.178±0.0050.178_± 0.005 0.588±0.0090.588_± 0.009 16.23±0.4616.23_± 0.46 0.167±0.0070.167_± 0.007 2.79±0.222.79_± 0.22 Gibberish 0.233±0.0050.233_± 0.005 0.672±0.0060.672_± 0.006 19.83±0.8219.83_± 0.82 0.184±0.0080.184_± 0.008 3.14±0.343.14_± 0.34 Baseline personas MPAQ 0.216±0.0020.216_± 0.002 0.652±0.0030.652_± 0.003 19.05±0.4919.05_± 0.49 0.186±0.0030.186_± 0.003 3.46±0.273.46_± 0.27 Individualized personas Random 0.202±0.0040.202_± 0.004 0.621±0.0070.621_± 0.007 12.59±0.6312.59_± 0.63 0.202±0.0040.202_± 0.004 1.54±0.121.54_± 0.12 Typical 0.197±0.0060.197_± 0.006 0.615±0.0100.615_± 0.010 12.78±0.2612.78_± 0.26 0.204±0.0030.204_± 0.003 1.63±0.121.63_± 0.12 Selected personas from base pool Coverage, cosine/L2 0.207±0.0060.207_± 0.006 0.629±0.0080.629_± 0.008 12.71±0.2812.71_± 0.28 0.203±0.0070.203_± 0.007 1.58±0.081.58_± 0.08 Coverage, Mahalanobis 0.193±0.0030.193_± 0.003 0.604±0.0050.604_± 0.005 11.77±0.2011.77_± 0.20 0.200±0.0070.200_± 0.007 1.50±0.061.50_± 0.06 Dispersion, cosine/L2 0.208±0.0030.208_± 0.003 0.633±0.0040.633_± 0.004 13.36±0.1813.36_± 0.18 0.213±0.0050.213_± 0.005 1.98±0.141.98_± 0.14 Dispersion, Mahalanobis 0.202±0.0040.202_± 0.004 0.622±0.0050.622_± 0.005 12.95±0.3512.95_± 0.35 0.210±0.0050.210_± 0.005 1.96±0.171.96_± 0.17 Generated personas (expanded pool size) MCMC (8,553) 0.207±0.0040.207_± 0.004 0.631±0.0060.631_± 0.006 14.04±0.5914.04_± 0.59 0.211±0.0010.211_± 0.001 2.21±0.232.21_± 0.23 Evolution (1,869) 0.214±0.0050.214_± 0.005 0.644±0.0070.644_± 0.007 15.47±0.3915.47_± 0.39 0.210±0.0050.210_± 0.005 2.53±0.202.53_± 0.20 AUT-Evolution (3,397) 0.216±0.0060.216_± 0.006 0.646±0.0090.646_± 0.009 15.66±0.2115.66_± 0.21 0.218±0.0070.218_± 0.007 2.98±0.162.98_± 0.16 Selected personas ++ Creativity-enhanced prompt Coverage, cosine/L2 0.190±0.0050.190_± 0.005 0.610±0.0090.610_± 0.009 20.44±0.3220.44_± 0.32 0.150±0.0060.150_± 0.006 5.17±0.215.17_± 0.21 Coverage, Mahalanobis 0.190±0.0030.190_± 0.003 0.609±0.0040.609_± 0.004 20.45±0.1720.45_± 0.17 0.154±0.0020.154_± 0.002 5.22±0.155.22_± 0.15 Dispersion, cosine/L2 0.195±0.0050.195_± 0.005 0.618±0.0090.618_± 0.009 20.57±0.2020.57_± 0.20 0.159±0.0070.159_± 0.007 6.43±0.226.43_± 0.22 Dispersion, Mahalanobis 0.191±0.0040.191_± 0.004 0.611±0.0080.611_± 0.008 20.57±0.5120.57_± 0.51 0.158±0.0070.158_± 0.007 6.73±0.296.73_± 0.29 Generated personas ++ Creativity-enhanced prompt MCMC 0.193±0.0040.193_± 0.004 0.615±0.0080.615_± 0.008 21.16±0.4121.16_± 0.41 0.156±0.0050.156_± 0.005 6.75±0.216.75_± 0.21 Evolution 0.197±0.0050.197_± 0.005 0.621±0.0090.621_± 0.009 21.20±0.2521.20_± 0.25 0.157±0.0050.157_± 0.005 6.11±0.226.11_± 0.22 AUT-Evolution 0.199±0.0050.199_± 0.005 0.625±0.0080.625_± 0.008 21.86±0.6521.86_± 0.65 0.156±0.0050.156_± 0.005 7.00±0.267.00_± 0.26 Table 13: Response diversity restricted to the three objects with human reference data (book, fork, tin can), so that all variants are directly comparable to Stevenson. Cells, bolding, and column conventions are as in Table 12. Mean pairwise distance Convex hull extent Variant Cosine L2 Mahal. Cosine/L2 Mahal. Standard persona Common-Use 0.182±0.0050.182_± 0.005 0.583±0.0100.583_± 0.010 15.49±0.6615.49_± 0.66 0.124±0.0090.124_± 0.009 2.16±0.382.16_± 0.38 Alternative-Use 0.195±0.0030.195_± 0.003 0.601±0.0050.601_± 0.005 11.38±0.3911.38_± 0.39 0.173±0.0120.173_± 0.012 1.20±0.131.20_± 0.13 Expert 0.185±0.0050.185_± 0.005 0.590±0.0050.590_± 0.005 13.76±0.3213.76_± 0.32 0.170±0.0120.170_± 0.012 2.01±0.262.01_± 0.26 Creativity-enhanced 0.174±0.0060.174_± 0.006 0.583±0.0090.583_± 0.009 19.54±0.3119.54_± 0.31 0.138±0.0060.138_± 0.006 5.06±0.405.06_± 0.40 ZS-CoT 0.162±0.0060.162_± 0.006 0.556±0.0110.556_± 0.011 13.62±0.4613.62_± 0.46 0.159±0.0140.159_± 0.014 1.99±0.141.99_± 0.14 Step-Back 0.180±0.0040.180_± 0.004 0.584±0.0090.584_± 0.009 12.79±0.2012.79_± 0.20 0.162±0.0120.162_± 0.012 1.77±0.181.77_± 0.18 DMAD 0.178±0.0090.178_± 0.009 0.585±0.0160.585_± 0.016 14.98±0.7714.98_± 0.77 0.168±0.0080.168_± 0.008 2.28±0.192.28_± 0.19 Gibberish 0.235±0.0090.235_± 0.009 0.675±0.0160.675_± 0.016 20.50±2.5820.50_± 2.58 0.178±0.0130.178_± 0.013 3.27±0.633.27_± 0.63 Baseline personas MPAQ 0.216±0.0070.216_± 0.007 0.651±0.0110.651_± 0.011 17.70±0.5917.70_± 0.59 0.198±0.0080.198_± 0.008 2.92±0.382.92_± 0.38 Individualized personas Random 0.206±0.0040.206_± 0.004 0.629±0.0070.629_± 0.007 12.52±0.6612.52_± 0.66 0.213±0.0080.213_± 0.008 1.48±0.151.48_± 0.15 Typical 0.206±0.0060.206_± 0.006 0.629±0.0100.629_± 0.010 12.70±0.3112.70_± 0.31 0.216±0.0130.216_± 0.013 1.58±0.211.58_± 0.21 Selected personas from base pool Coverage, cosine/L2 0.214±0.0080.214_± 0.008 0.639±0.0130.639_± 0.013 12.66±0.6312.66_± 0.63 0.214±0.0080.214_± 0.008 1.57±0.161.57_± 0.16 Coverage, Mahalanobis 0.205±0.0080.205_± 0.008 0.624±0.0160.624_± 0.016 11.90±0.4911.90_± 0.49 0.214±0.0090.214_± 0.009 1.52±0.151.52_± 0.15 Dispersion, cosine/L2 0.209±0.0020.209_± 0.002 0.638±0.0030.638_± 0.003 13.11±0.4513.11_± 0.45 0.220±0.0050.220_± 0.005 1.87±0.331.87_± 0.33 Dispersion, Mahalanobis 0.208±0.0030.208_± 0.003 0.634±0.0040.634_± 0.004 12.87±0.3912.87_± 0.39 0.224±0.0030.224_± 0.003 1.95±0.321.95_± 0.32 Generated personas (expanded pool size) MCMC (8,553) 0.211±0.0090.211_± 0.009 0.639±0.0140.639_± 0.014 13.92±1.3513.92_± 1.35 0.223±0.0070.223_± 0.007 2.21±0.482.21_± 0.48 Evolution (1,869) 0.214±0.0080.214_± 0.008 0.646±0.0120.646_± 0.012 15.41±0.6115.41_± 0.61 0.212±0.0030.212_± 0.003 2.56±0.412.56_± 0.41 AUT-Evolution (3,397) 0.222±0.0060.222_± 0.006 0.658±0.0110.658_± 0.011 16.03±0.5616.03_± 0.56 0.227±0.0070.227_± 0.007 3.08±0.343.08_± 0.34 Selected personas ++ Creativity-enhanced prompt Coverage, cosine/L2 0.186±0.0050.186_± 0.005 0.602±0.0090.602_± 0.009 19.84±0.6019.84_± 0.60 0.145±0.0060.145_± 0.006 5.08±0.325.08_± 0.32 Coverage, Mahalanobis 0.186±0.0050.186_± 0.005 0.604±0.0080.604_± 0.008 19.94±0.3919.94_± 0.39 0.144±0.0050.144_± 0.005 5.22±0.255.22_± 0.25 Dispersion, cosine/L2 0.190±0.0050.190_± 0.005 0.608±0.0100.608_± 0.010 19.85±0.4619.85_± 0.46 0.153±0.0100.153_± 0.010 6.17±0.526.17_± 0.52 Dispersion, Mahalanobis 0.184±0.0020.184_± 0.002 0.600±0.0040.600_± 0.004 20.18±0.6020.18_± 0.60 0.148±0.0050.148_± 0.005 6.59±0.386.59_± 0.38 Generated personas ++ Creativity-enhanced prompt MCMC 0.186±0.0030.186_± 0.003 0.604±0.0060.604_± 0.006 20.62±0.2420.62_± 0.24 0.147±0.0040.147_± 0.004 6.58±0.286.58_± 0.28 Evolution 0.190±0.0050.190_± 0.005 0.610±0.0090.610_± 0.009 20.66±0.7320.66_± 0.73 0.148±0.0110.148_± 0.011 5.98±0.495.98_± 0.49 AUT-Evolution 0.196±0.0060.196_± 0.006 0.620±0.0110.620_± 0.011 21.48±0.7521.48_± 0.75 0.148±0.0030.148_± 0.003 6.92±0.406.92_± 0.40 Human Stevenson 0.149±0.0090.149_± 0.009 0.539±0.0150.539_± 0.015 17.80±0.9317.80_± 0.93 0.101±0.0060.101_± 0.006 2.26±0.342.26_± 0.34 Table 14: Response originality across all seven AUT objects, measured both as distance to the prompted object and as directed distance to the common-use set, following equation 57. Cells give the mean over valid uses, averaged over objects; higher is more original. Common-Use defines the common-use reference set CoC_o, so its distance to that set is zero by construction and is omitted (—). Subscripts denote the 95% confidence interval across 5 independent random seeds. The final two groups re-run our selection and generation methods on the Creativity-enhanced prompt, showing that persona diversification composes with a stronger prompting strategy. Bold marks the best value per column within the standard, selected and generated groups separately, with tied values bolded jointly; baseline and individualized personas are not bolded. Generated variants report their expanded pool size in parentheses. To object To common uses Variant Cosine L2 Mahal. Cosine L2 Mahal. Standard persona Common-Use 0.142±0.0020.142_± 0.002 0.525±0.0040.525_± 0.004 15.28±0.1615.28_± 0.16 — — — Alternative-Use 0.170±0.0010.170_± 0.001 0.580±0.0020.580_± 0.002 13.60±0.1413.60_± 0.14 0.114±0.0080.114_± 0.008 0.439±0.0190.439_± 0.019 10.79±0.3210.79_± 0.32 Expert 0.164±0.0020.164_± 0.002 0.568±0.0040.568_± 0.004 14.72±0.1714.72_± 0.17 0.104±0.0030.104_± 0.003 0.417±0.0100.417_± 0.010 11.25±0.4011.25_± 0.40 Creativity-enhanced 0.173±0.0040.173_± 0.004 0.586±0.0070.586_± 0.007 18.13±0.3518.13_± 0.35 0.174±0.0060.174_± 0.006 0.585±0.0110.585_± 0.011 17.15±0.4017.15_± 0.40 ZS-CoT 0.146±0.0020.146_± 0.002 0.536±0.0040.536_± 0.004 14.67±0.1014.67_± 0.10 0.125±0.0030.125_± 0.003 0.483±0.0070.483_± 0.007 13.02±0.4313.02_± 0.43 Step-Back 0.165±0.0020.165_± 0.002 0.571±0.0030.571_± 0.003 14.91±0.1814.91_± 0.18 0.129±0.0020.129_± 0.002 0.488±0.0070.488_± 0.007 12.86±0.2012.86_± 0.20 DMAD 0.151±0.0040.151_± 0.004 0.547±0.0070.547_± 0.007 15.19±0.1915.19_± 0.19 0.135±0.0060.135_± 0.006 0.504±0.0170.504_± 0.017 13.90±0.4913.90_± 0.49 Gibberish 0.193±0.0020.193_± 0.002 0.618±0.0040.618_± 0.004 17.72±0.5217.72_± 0.52 0.156±0.0090.156_± 0.009 0.541±0.0180.541_± 0.018 15.89±0.6815.89_± 0.68 Baseline personas MPAQ 0.180±0.0010.180_± 0.001 0.596±0.0020.596_± 0.002 16.95±0.2816.95_± 0.28 0.162±0.0020.162_± 0.002 0.560±0.0040.560_± 0.004 15.46±0.3415.46_± 0.34 Individualized personas Random 0.184±0.0040.184_± 0.004 0.603±0.0070.603_± 0.007 13.90±0.3113.90_± 0.31 0.143±0.0070.143_± 0.007 0.516±0.0150.516_± 0.015 11.89±0.4811.89_± 0.48 Typical 0.179±0.0030.179_± 0.003 0.595±0.0050.595_± 0.005 13.91±0.1313.91_± 0.13 0.139±0.0050.139_± 0.005 0.509±0.0110.509_± 0.011 11.96±0.3011.96_± 0.30 Selected personas from base pool Coverage, cosine 0.188±0.0050.188_± 0.005 0.609±0.0080.609_± 0.008 13.91±0.1413.91_± 0.14 0.145±0.0040.145_± 0.004 0.520±0.0080.520_± 0.008 12.03±0.3712.03_± 0.37 Coverage, Mahalanobis 0.179±0.0010.179_± 0.001 0.594±0.0020.594_± 0.002 13.51±0.1113.51_± 0.11 0.133±0.0060.133_± 0.006 0.495±0.0120.495_± 0.012 11.46±0.2811.46_± 0.28 Dispersion, cosine 0.184±0.0020.184_± 0.002 0.603±0.0030.603_± 0.003 14.20±0.0514.20_± 0.05 0.146±0.0040.146_± 0.004 0.524±0.0100.524_± 0.010 12.34±0.2712.34_± 0.27 Dispersion, Mahalanobis 0.181±0.0030.181_± 0.003 0.598±0.0040.598_± 0.004 13.98±0.1813.98_± 0.18 0.138±0.0050.138_± 0.005 0.506±0.0090.506_± 0.009 12.00±0.2812.00_± 0.28 Generated personas (expanded pool size) MCMC (8,553) 0.182±0.0020.182_± 0.002 0.600±0.0030.600_± 0.003 14.51±0.3014.51_± 0.30 0.144±0.0050.144_± 0.005 0.519±0.0140.519_± 0.014 12.66±0.5512.66_± 0.55 Evolution (1,869) 0.190±0.0040.190_± 0.004 0.612±0.0070.612_± 0.007 15.22±0.1915.22_± 0.19 0.162±0.0030.162_± 0.003 0.556±0.0060.556_± 0.006 13.61±0.1713.61_± 0.17 AUT-Evolution (3,397) 0.188±0.0020.188_± 0.002 0.610±0.0030.610_± 0.003 15.29±0.1215.29_± 0.12 0.156±0.0040.156_± 0.004 0.544±0.0060.544_± 0.006 13.58±0.1113.58_± 0.11 Selected personas ++ Creativity-enhanced prompt Coverage, cosine 0.187±0.0030.187_± 0.003 0.609±0.0050.609_± 0.005 18.45±0.2018.45_± 0.20 0.181±0.0030.181_± 0.003 0.598±0.0050.598_± 0.005 17.25±0.1217.25_± 0.12 Coverage, Mahalanobis 0.184±0.0030.184_± 0.003 0.603±0.0050.603_± 0.005 18.33±0.1118.33_± 0.11 0.177±0.0040.177_± 0.004 0.590±0.0060.590_± 0.006 17.20±0.1517.20_± 0.15 Dispersion, cosine 0.185±0.0050.185_± 0.005 0.604±0.0080.604_± 0.008 18.31±0.2718.31_± 0.27 0.179±0.0070.179_± 0.007 0.593±0.0130.593_± 0.013 17.24±0.2617.24_± 0.26 Dispersion, Mahalanobis 0.182±0.0050.182_± 0.005 0.600±0.0080.600_± 0.008 18.36±0.3618.36_± 0.36 0.178±0.0070.178_± 0.007 0.592±0.0110.592_± 0.011 17.22±0.3917.22_± 0.39 Generated personas ++ Creativity-enhanced prompt MCMC 0.183±0.0050.183_± 0.005 0.602±0.0090.602_± 0.009 18.71±0.2218.71_± 0.22 0.182±0.0060.182_± 0.006 0.598±0.0090.598_± 0.009 17.64±0.2117.64_± 0.21 Evolution 0.187±0.0050.187_± 0.005 0.608±0.0080.608_± 0.008 18.80±0.1718.80_± 0.17 0.183±0.0060.183_± 0.006 0.600±0.0110.600_± 0.011 17.65±0.1417.65_± 0.14 AUT-Evolution 0.187±0.0060.187_± 0.006 0.608±0.0090.608_± 0.009 19.14±0.4119.14_± 0.41 0.186±0.0090.186_± 0.009 0.605±0.0140.605_± 0.014 18.02±0.3918.02_± 0.39 Table 15: Response originality restricted to the three objects with human reference data (book, fork, tin can), so that all variants are directly comparable to Stevenson. Cells, bolding, and column conventions are as in Table 14. Common-Use defines the common-use reference set CoC_o, so its distance to that set is zero by construction and is omitted (—). To object To common uses Variant Cosine L2 Mahal. Cosine L2 Mahal. Standard persona Common-Use 0.182±0.0050.182_± 0.005 0.583±0.0100.583_± 0.010 15.49±0.6615.49_± 0.66 — — — Alternative-Use 0.195±0.0030.195_± 0.003 0.601±0.0050.601_± 0.005 11.38±0.3911.38_± 0.39 0.173±0.0120.173_± 0.012 1.20±0.131.20_± 0.13 10.79±0.3210.79_± 0.32 Expert 0.185±0.0050.185_± 0.005 0.590±0.0050.590_± 0.005 13.76±0.3213.76_± 0.32 0.170±0.0120.170_± 0.012 2.01±0.262.01_± 0.26 11.25±0.4011.25_± 0.40 Creativity-enhanced 0.174±0.0060.174_± 0.006 0.583±0.0090.583_± 0.009 19.54±0.3119.54_± 0.31 0.138±0.0060.138_± 0.006 5.06±0.405.06_± 0.40 17.15±0.4017.15_± 0.40 ZS-CoT 0.162±0.0060.162_± 0.006 0.556±0.0110.556_± 0.011 13.62±0.4613.62_± 0.46 0.159±0.0140.159_± 0.014 1.99±0.141.99_± 0.14 13.02±0.4313.02_± 0.43 Step-Back 0.180±0.0040.180_± 0.004 0.584±0.0090.584_± 0.009 12.79±0.2012.79_± 0.20 0.162±0.0120.162_± 0.012 1.77±0.181.77_± 0.18 12.86±0.2012.86_± 0.20 DMAD 0.178±0.0090.178_± 0.009 0.585±0.0160.585_± 0.016 14.98±0.7714.98_± 0.77 0.168±0.0080.168_± 0.008 2.28±0.192.28_± 0.19 13.90±0.4913.90_± 0.49 Gibberish 0.235±0.0090.235_± 0.009 0.675±0.0160.675_± 0.016 20.50±2.5820.50_± 2.58 0.178±0.0130.178_± 0.013 3.27±0.633.27_± 0.63 15.89±0.6815.89_± 0.68 Baseline personas MPAQ 0.216±0.0070.216_± 0.007 0.651±0.0110.651_± 0.011 17.70±0.5917.70_± 0.59 0.198±0.0080.198_± 0.008 2.92±0.382.92_± 0.38 15.46±0.3415.46_± 0.34 Individualized personas Random 0.206±0.0040.206_± 0.004 0.629±0.0070.629_± 0.007 12.52±0.6612.52_± 0.66 0.213±0.0080.213_± 0.008 1.48±0.151.48_± 0.15 11.89±0.4811.89_± 0.48 Typical 0.206±0.0060.206_± 0.006 0.629±0.0100.629_± 0.010 12.70±0.3112.70_± 0.31 0.216±0.0130.216_± 0.013 1.58±0.211.58_± 0.21 11.96±0.3011.96_± 0.30 Selected personas from base pool Coverage, cosine 0.214±0.0080.214_± 0.008 0.639±0.0130.639_± 0.013 12.66±0.6312.66_± 0.63 0.214±0.0080.214_± 0.008 1.57±0.161.57_± 0.16 12.03±0.3712.03_± 0.37 Coverage, Mahalanobis 0.205±0.0080.205_± 0.008 0.624±0.0160.624_± 0.016 11.90±0.4911.90_± 0.49 0.214±0.0090.214_± 0.009 1.52±0.151.52_± 0.15 11.46±0.2811.46_± 0.28 Dispersion, cosine 0.209±0.0020.209_± 0.002 0.638±0.0030.638_± 0.003 13.11±0.4513.11_± 0.45 0.220±0.0050.220_± 0.005 1.87±0.331.87_± 0.33 12.34±0.2712.34_± 0.27 Dispersion, Mahalanobis 0.208±0.0030.208_± 0.003 0.634±0.0040.634_± 0.004 12.87±0.3912.87_± 0.39 0.224±0.0030.224_± 0.003 1.95±0.321.95_± 0.32 12.00±0.2812.00_± 0.28 Generated personas (expanded pool size) MCMC (8,553) 0.211±0.0090.211_± 0.009 0.639±0.0140.639_± 0.014 13.92±1.3513.92_± 1.35 0.223±0.0070.223_± 0.007 2.21±0.482.21_± 0.48 12.66±0.5512.66_± 0.55 Evolution (1,869) 0.214±0.0080.214_± 0.008 0.646±0.0120.646_± 0.012 15.41±0.6115.41_± 0.61 0.212±0.0030.212_± 0.003 2.56±0.412.56_± 0.41 13.61±0.1713.61_± 0.17 AUT-Evolution (3,397) 0.222±0.0060.222_± 0.006 0.658±0.0110.658_± 0.011 16.03±0.5616.03_± 0.56 0.227±0.0070.227_± 0.007 3.08±0.343.08_± 0.34 13.58±0.1113.58_± 0.11 Selected personas ++ Creativity-enhanced prompt Coverage, cosine 0.186±0.0050.186_± 0.005 0.602±0.0090.602_± 0.009 19.84±0.6019.84_± 0.60 0.145±0.0060.145_± 0.006 5.08±0.325.08_± 0.32 17.25±0.1217.25_± 0.12 Coverage, Mahalanobis 0.186±0.0050.186_± 0.005 0.604±0.0080.604_± 0.008 19.94±0.3919.94_± 0.39 0.144±0.0050.144_± 0.005 5.22±0.255.22_± 0.25 17.20±0.1517.20_± 0.15 Dispersion, cosine 0.190±0.0050.190_± 0.005 0.608±0.0100.608_± 0.010 19.85±0.4619.85_± 0.46 0.153±0.0100.153_± 0.010 6.17±0.526.17_± 0.52 17.24±0.2617.24_± 0.26 Dispersion, Mahalanobis 0.184±0.0020.184_± 0.002 0.600±0.0040.600_± 0.004 20.18±0.6020.18_± 0.60 0.148±0.0050.148_± 0.005 6.59±0.386.59_± 0.38 17.22±0.3917.22_± 0.39 Generated personas ++ Creativity-enhanced prompt MCMC 0.186±0.0030.186_± 0.003 0.604±0.0060.604_± 0.006 20.62±0.2420.62_± 0.24 0.147±0.0040.147_± 0.004 6.58±0.286.58_± 0.28 17.64±0.2117.64_± 0.21 Evolution 0.190±0.0050.190_± 0.005 0.610±0.0090.610_± 0.009 20.66±0.7320.66_± 0.73 0.148±0.0110.148_± 0.011 5.98±0.495.98_± 0.49 17.65±0.1417.65_± 0.14 AUT-Evolution 0.196±0.0060.196_± 0.006 0.620±0.0110.620_± 0.011 21.48±0.7521.48_± 0.75 0.148±0.0030.148_± 0.003 6.92±0.406.92_± 0.40 18.02±0.3918.02_± 0.39 Human Stevenson 0.119±0.0000.119_± 0.000 0.480±0.0000.480_± 0.000 15.51±0.0015.51_± 0.00 0.144±0.0000.144_± 0.000 0.533±0.0000.533_± 0.000 14.15±0.0014.15_± 0.00 Table 16: Response flexibility, reported over all seven AUT objects and, for comparability with Stevenson, over the three objects with human reference data (175175 and 7575 generated uses per variant, respectively). Both measures are size-controlled: Flexibility is the number of distinct use categories and Vendi is the Vendi score equation 53, each divided by the number of valid uses, so that variants admitting different numbers of uses remain comparable. Cells give the mean over objects. Subscripts denote the 95% confidence interval across 5 independent random seeds; higher is more flexible. Stevenson exists only for the human-reference objects, so its seven-object cells are omitted (—). The final two groups re-run our selection and generation methods on the Creativity-enhanced prompt, showing that persona diversification composes with a stronger prompting strategy. Bold marks the best value per column within the standard, selected and generated groups separately, with tied values bolded jointly; baseline, individualized and human rows are not bolded. Generated variants report their expanded pool size in parentheses. All seven objects Human-reference objects Variant Flexibility Vendi Flexibility Vendi Standard persona Common-Use 0.160±0.0110.160_± 0.011 0.086±0.0030.086_± 0.003 0.214±0.0250.214_± 0.025 0.094±0.0040.094_± 0.004 Alternative-Use 0.288±0.0220.288_± 0.022 0.096±0.0020.096_± 0.002 0.284±0.0120.284_± 0.012 0.099±0.0030.099_± 0.003 Expert 0.300±0.0260.300_± 0.026 0.094±0.0010.094_± 0.001 0.320±0.0170.320_± 0.017 0.096±0.0010.096_± 0.001 Creativity-enhanced 0.503±0.0560.503_± 0.056 0.104±0.0020.104_± 0.002 0.467±0.0710.467_± 0.071 0.101±0.0020.101_± 0.002 ZS-CoT 0.332±0.0250.332_± 0.025 0.094±0.0010.094_± 0.001 0.323±0.0440.323_± 0.044 0.091±0.0020.091_± 0.002 Step-Back 0.303±0.0220.303_± 0.022 0.096±0.0040.096_± 0.004 0.280±0.0320.280_± 0.032 0.097±0.0030.097_± 0.003 DMAD 0.372±0.0250.372_± 0.025 0.102±0.0030.102_± 0.003 0.350±0.0340.350_± 0.034 0.098±0.0040.098_± 0.004 Gibberish 0.384±0.0610.384_± 0.061 0.151±0.0130.151_± 0.013 0.359±0.0310.359_± 0.031 0.158±0.0270.158_± 0.027 Baseline personas MPAQ 0.561±0.0300.561_± 0.030 0.130±0.0060.130_± 0.006 0.491±0.0570.491_± 0.057 0.129±0.0060.129_± 0.006 Individualized personas Random 0.331±0.0240.331_± 0.024 0.107±0.0030.107_± 0.003 0.333±0.0310.333_± 0.031 0.109±0.0030.109_± 0.003 Typical 0.330±0.0160.330_± 0.016 0.105±0.0020.105_± 0.002 0.323±0.0220.323_± 0.022 0.108±0.0030.108_± 0.003 Selected personas from base pool Coverage, cosine 0.338±0.0250.338_± 0.025 0.110±0.0030.110_± 0.003 0.323±0.0250.323_± 0.025 0.111±0.0040.111_± 0.004 Coverage, Mahalanobis 0.337±0.0110.337_± 0.011 0.101±0.0020.101_± 0.002 0.329±0.0410.329_± 0.041 0.105±0.0050.105_± 0.005 Dispersion, cosine 0.379±0.0260.379_± 0.026 0.112±0.0010.112_± 0.001 0.356±0.0480.356_± 0.048 0.113±0.0010.113_± 0.001 Dispersion, Mahalanobis 0.357±0.0480.357_± 0.048 0.108±0.0020.108_± 0.002 0.353±0.0510.353_± 0.051 0.111±0.0010.111_± 0.001 Generated personas (expanded pool size) MCMC (8,553) 0.378±0.0520.378_± 0.052 0.112±0.0010.112_± 0.001 0.348±0.0290.348_± 0.029 0.113±0.0050.113_± 0.005 Evolution (1,869) 0.431±0.0450.431_± 0.045 0.117±0.0020.117_± 0.002 0.411±0.1020.411_± 0.102 0.117±0.0030.117_± 0.003 AUT-Evolution (3,397) 0.463±0.0520.463_± 0.052 0.118±0.0020.118_± 0.002 0.490±0.0450.490_± 0.045 0.121±0.0040.121_± 0.004 Selected personas ++ Creativity-enhanced prompt Coverage, cosine 0.508±0.0820.508_± 0.082 0.109±0.0030.109_± 0.003 0.463±0.1010.463_± 0.101 0.107±0.0020.107_± 0.002 Coverage, Mahalanobis 0.518±0.0660.518_± 0.066 0.108±0.0010.108_± 0.001 0.496±0.0900.496_± 0.090 0.107±0.0030.107_± 0.003 Dispersion, cosine 0.543±0.0520.543_± 0.052 0.111±0.0030.111_± 0.003 0.514±0.0940.514_± 0.094 0.109±0.0030.109_± 0.003 Dispersion, Mahalanobis 0.560±0.0510.560_± 0.051 0.109±0.0030.109_± 0.003 0.526±0.1170.526_± 0.117 0.106±0.0010.106_± 0.001 Generated personas ++ Creativity-enhanced prompt MCMC 0.560±0.0450.560_± 0.045 0.111±0.0020.111_± 0.002 0.526±0.1080.526_± 0.108 0.108±0.0020.108_± 0.002 Evolution 0.580±0.0460.580_± 0.046 0.112±0.0030.112_± 0.003 0.562±0.0620.562_± 0.062 0.110±0.0040.110_± 0.004 AUT-Evolution 0.612±0.0260.612_± 0.026 0.114±0.0030.114_± 0.003 0.606±0.1230.606_± 0.123 0.113±0.0030.113_± 0.003 Human Stevenson — — 0.739±0.0430.739_± 0.043 0.104±0.0010.104_± 0.001 Table 17: Response surprise, reported over all seven AUT objects and, for comparability with Stevenson, over the three objects with human reference data (175175 and 7575 generated uses per variant, respectively). Both measures are size-controlled by the number of valid uses: Surprise is the proportion of uses judged surprising relative to the common-use reference set, and Wtd. Vendi is the novelty-weighted Vendi score equation 53 under the same normalization. Common-Use defines the reference set against which surprise is measured, so its surprise is zero by construction and is omitted (—); Stevenson exists only for the human-reference objects. Cells give the mean over objects. Subscripts denote the 95% confidence interval across 5 independent random seeds; higher is more surprising. The final two groups re-run our selection and generation methods on the Creativity-enhanced prompt, showing that persona diversification composes with a stronger prompting strategy. Bolding conventions are as in Table 16. All seven objects Human-reference objects Variant Surprise Wtd. Vendi Surprise Wtd. Vendi Standard persona Common-Use — — — — Alternative-Use 0.506±0.0620.506_± 0.062 0.094±0.0020.094_± 0.002 0.397±0.0870.397_± 0.087 0.093±0.0040.093_± 0.004 Expert 0.415±0.0200.415_± 0.020 0.093±0.0020.093_± 0.002 0.317±0.0140.317_± 0.014 0.098±0.0020.098_± 0.002 Creativity-enhanced 0.837±0.0670.837_± 0.067 0.105±0.0020.105_± 0.002 0.717±0.1310.717_± 0.131 0.102±0.0020.102_± 0.002 ZS-CoT 0.556±0.0640.556_± 0.064 0.094±0.0020.094_± 0.002 0.390±0.0630.390_± 0.063 0.089±0.0030.089_± 0.003 Step-Back 0.558±0.0680.558_± 0.068 0.095±0.0050.095_± 0.005 0.413±0.0870.413_± 0.087 0.095±0.0050.095_± 0.005 DMAD 0.567±0.0570.567_± 0.057 0.103±0.0040.103_± 0.004 0.382±0.1310.382_± 0.131 0.100±0.0050.100_± 0.005 Gibberish 0.525±0.0580.525_± 0.058 0.150±0.0120.150_± 0.012 0.398±0.0260.398_± 0.026 0.156±0.0250.156_± 0.025 Baseline personas MPAQ 0.686±0.0550.686_± 0.055 0.132±0.0060.132_± 0.006 0.535±0.0900.535_± 0.090 0.131±0.0060.131_± 0.006 Individualized personas Random 0.588±0.0510.588_± 0.051 0.107±0.0030.107_± 0.003 0.467±0.0610.467_± 0.061 0.109±0.0040.109_± 0.004 Typical 0.572±0.0510.572_± 0.051 0.105±0.0030.105_± 0.003 0.454±0.0550.454_± 0.055 0.107±0.0030.107_± 0.003 Selected personas from base pool Coverage, cosine 0.574±0.0570.574_± 0.057 0.110±0.0030.110_± 0.003 0.445±0.0970.445_± 0.097 0.111±0.0050.111_± 0.005 Coverage, Mahalanobis 0.548±0.0480.548_± 0.048 0.101±0.0030.101_± 0.003 0.385±0.0290.385_± 0.029 0.104±0.0050.104_± 0.005 Dispersion, cosine 0.588±0.0810.588_± 0.081 0.112±0.0010.112_± 0.001 0.449±0.0860.449_± 0.086 0.111±0.0010.111_± 0.001 Dispersion, Mahalanobis 0.567±0.0750.567_± 0.075 0.108±0.0020.108_± 0.002 0.406±0.0690.406_± 0.069 0.110±0.0030.110_± 0.003 Generated personas (expanded pool size) MCMC (8,553) 0.566±0.0780.566_± 0.078 0.113±0.0020.113_± 0.002 0.411±0.0780.411_± 0.078 0.113±0.0070.113_± 0.007 Evolution (1,869) 0.637±0.0640.637_± 0.064 0.118±0.0020.118_± 0.002 0.514±0.1250.514_± 0.125 0.118±0.0040.118_± 0.004 AUT-Evolution (3,397) 0.637±0.0680.637_± 0.068 0.118±0.0020.118_± 0.002 0.522±0.0660.522_± 0.066 0.120±0.0050.120_± 0.005 Selected personas ++ Creativity-enhanced prompt Coverage, cosine 0.843±0.0610.843_± 0.061 0.110±0.0030.110_± 0.003 0.739±0.1260.739_± 0.126 0.108±0.0020.108_± 0.002 Coverage, Mahalanobis 0.815±0.0740.815_± 0.074 0.109±0.0010.109_± 0.001 0.707±0.1300.707_± 0.130 0.108±0.0020.108_± 0.002 Dispersion, cosine 0.833±0.0430.833_± 0.043 0.113±0.0030.113_± 0.003 0.763±0.0710.763_± 0.071 0.110±0.0030.110_± 0.003 Dispersion, Mahalanobis 0.839±0.0320.839_± 0.032 0.111±0.0030.111_± 0.003 0.781±0.0790.781_± 0.079 0.108±0.0010.108_± 0.001 Generated personas ++ Creativity-enhanced prompt MCMC 0.847±0.0210.847_± 0.021 0.112±0.0020.112_± 0.002 0.792±0.0670.792_± 0.067 0.109±0.0020.109_± 0.002 Evolution 0.845±0.0430.845_± 0.043 0.113±0.0030.113_± 0.003 0.781±0.0790.781_± 0.079 0.111±0.0040.111_± 0.004 AUT-Evolution 0.872±0.0390.872_± 0.039 0.115±0.0030.115_± 0.003 0.822±0.0890.822_± 0.089 0.115±0.0030.115_± 0.003 Human Stevenson — — 0.741±0.0650.741_± 0.065 0.106±0.0010.106_± 0.001 Table 18: LLM-judged creativity on the AUT across all seven objects. A Qwen3.6-27B judge, independent of the Gemma-4-31B generator, rates every use for Originality, Surprise and Utility on a 11–55 scale following Stevenson et al. (2022), and for holistic creativity following Góes et al. (2023); we report the rank-normalized form of that holistic judgment, which correlates with its raw form at r=0.996r=0.996. Cells give the mean over judged uses (at most 175175 per variant). The judged set is every use not rejected outright by the validity gate of Section D.2, admitted and uncertain uses, so it is slightly larger than the admitted set of Table 11. Subscripts denote the 95% confidence interval across 5 independent random seeds. The final two groups re-run our selection and generation methods on the Creativity-enhanced prompt, showing that persona diversification composes with a stronger prompting strategy. Originality, Surprise and Creativity are target metrics, bolded per column within the standard, selected and generated groups separately (and within each creativity-enhanced group), with tied values bolded jointly. Utility is a control: conventional uses are the most useful by construction, so it necessarily peaks at Common-Use and is never bolded. Baseline, individualized and human rows are not bolded. Generated variants report their expanded pool size in parentheses. Stevenson et al. (2022) Góes et al. (2023) Variant Originality Surprise Utility Creativity (rank) Standard persona Common-Use 1.68±0.031.68_± 0.03 1.34±0.011.34_± 0.01 4.86±0.044.86_± 0.04 2.02±0.082.02_± 0.08 Alternative-Use 2.63±0.052.63_± 0.05 2.23±0.062.23_± 0.06 4.30±0.054.30_± 0.05 2.72±0.092.72_± 0.09 Expert 2.28±0.052.28_± 0.05 2.01±0.052.01_± 0.05 4.39±0.054.39_± 0.05 2.71±0.052.71_± 0.05 Creativity-enhanced 3.17±0.063.17_± 0.06 2.66±0.062.66_± 0.06 3.38±0.073.38_± 0.07 3.19±0.093.19_± 0.09 ZS-CoT 2.60±0.092.60_± 0.09 2.11±0.072.11_± 0.07 4.27±0.074.27_± 0.07 2.85±0.062.85_± 0.06 Step-Back 2.74±0.022.74_± 0.02 2.30±0.072.30_± 0.07 4.30±0.044.30_± 0.04 2.95±0.062.95_± 0.06 DMAD 2.65±0.112.65_± 0.11 2.17±0.092.17_± 0.09 4.17±0.114.17_± 0.11 2.89±0.072.89_± 0.07 Gibberish 2.85±0.062.85_± 0.06 2.33±0.072.33_± 0.07 4.02±0.104.02_± 0.10 2.57±0.092.57_± 0.09 Baseline personas MPAQ 3.29±0.043.29_± 0.04 2.65±0.062.65_± 0.06 3.59±0.043.59_± 0.04 3.05±0.073.05_± 0.07 Individualized personas Random 2.78±0.032.78_± 0.03 2.36±0.032.36_± 0.03 4.28±0.024.28_± 0.02 2.96±0.042.96_± 0.04 Typical 2.75±0.042.75_± 0.04 2.36±0.042.36_± 0.04 4.24±0.054.24_± 0.05 2.99±0.072.99_± 0.07 Selected personas from base pool Coverage, cosine 2.74±0.042.74_± 0.04 2.40±0.022.40_± 0.02 4.30±0.024.30_± 0.02 2.91±0.062.91_± 0.06 Coverage, Mahalanobis 2.66±0.022.66_± 0.02 2.29±0.052.29_± 0.05 4.27±0.044.27_± 0.04 2.86±0.082.86_± 0.08 Dispersion, cosine 2.78±0.062.78_± 0.06 2.40±0.042.40_± 0.04 4.27±0.094.27_± 0.09 2.91±0.062.91_± 0.06 Dispersion, Mahalanobis 2.74±0.082.74_± 0.08 2.36±0.062.36_± 0.06 4.23±0.094.23_± 0.09 2.92±0.042.92_± 0.04 Generated personas (expanded pool size) MCMC (8,553) 2.74±0.052.74_± 0.05 2.40±0.072.40_± 0.07 4.28±0.064.28_± 0.06 2.91±0.042.91_± 0.04 Evolution (1,869) 2.97±0.072.97_± 0.07 2.54±0.062.54_± 0.06 4.08±0.064.08_± 0.06 3.10±0.033.10_± 0.03 AUT-Evolution (3,397) 2.99±0.092.99_± 0.09 2.63±0.092.63_± 0.09 4.04±0.104.04_± 0.10 2.99±0.062.99_± 0.06 Selected personas ++ Creativity-enhanced prompt Coverage, cosine 3.22±0.073.22_± 0.07 2.75±0.042.75_± 0.04 3.43±0.103.43_± 0.10 3.25±0.043.25_± 0.04 Coverage, Mahalanobis 3.15±0.053.15_± 0.05 2.71±0.052.71_± 0.05 3.58±0.103.58_± 0.10 3.24±0.073.24_± 0.07 Dispersion, cosine 3.20±0.093.20_± 0.09 2.72±0.122.72_± 0.12 3.54±0.113.54_± 0.11 3.25±0.073.25_± 0.07 Dispersion, Mahalanobis 3.19±0.123.19_± 0.12 2.74±0.102.74_± 0.10 3.44±0.093.44_± 0.09 3.27±0.063.27_± 0.06 Generated personas ++ Creativity-enhanced prompt MCMC 3.24±0.153.24_± 0.15 2.76±0.102.76_± 0.10 3.39±0.103.39_± 0.10 3.31±0.053.31_± 0.05 Evolution 3.31±0.123.31_± 0.12 2.87±0.112.87_± 0.11 3.38±0.163.38_± 0.16 3.39±0.073.39_± 0.07 AUT-Evolution 3.38±0.133.38_± 0.13 2.94±0.132.94_± 0.13 3.33±0.143.33_± 0.14 3.32±0.063.32_± 0.06 Human Stevenson 2.78±0.002.78_± 0.00 1.77±0.001.77_± 0.00 3.49±0.033.49_± 0.03 2.07±0.032.07_± 0.03 Table 19: LLM-judged creativity restricted to the three objects with human reference data (book, fork, tin can), so that all variants are directly comparable to Stevenson (at most 7575 judged uses per variant). Scales, judge, bolding and control conventions are as in Table 18. Stevenson et al. (2022) Góes et al. (2023) Variant Originality Surprise Utility Creativity (rank) Standard persona Common-Use 1.84±0.041.84_± 0.04 1.45±0.031.45_± 0.03 4.84±0.034.84_± 0.03 2.29±0.122.29_± 0.12 Alternative-Use 2.37±0.092.37_± 0.09 2.07±0.052.07_± 0.05 4.39±0.084.39_± 0.08 2.78±0.162.78_± 0.16 Expert 2.18±0.042.18_± 0.04 2.00±0.062.00_± 0.06 4.40±0.084.40_± 0.08 2.83±0.072.83_± 0.07 Creativity-enhanced 3.05±0.183.05_± 0.18 2.66±0.062.66_± 0.06 3.52±0.113.52_± 0.11 3.19±0.093.19_± 0.09 ZS-CoT 2.37±0.112.37_± 0.11 1.96±0.121.96_± 0.12 4.49±0.104.49_± 0.10 2.88±0.122.88_± 0.12 Step-Back 2.52±0.072.52_± 0.07 2.16±0.072.16_± 0.07 4.31±0.094.31_± 0.09 2.95±0.072.95_± 0.07 DMAD 2.39±0.112.39_± 0.11 2.06±0.082.06_± 0.08 4.38±0.084.38_± 0.08 2.90±0.122.90_± 0.12 Gibberish 2.59±0.072.59_± 0.07 2.13±0.092.13_± 0.09 4.19±0.154.19_± 0.15 2.37±0.222.37_± 0.22 Baseline personas MPAQ 3.05±0.133.05_± 0.13 2.44±0.082.44_± 0.08 3.78±0.073.78_± 0.07 3.00±0.103.00_± 0.10 Individualized personas Random 2.52±0.052.52_± 0.05 2.18±0.042.18_± 0.04 4.35±0.034.35_± 0.03 3.03±0.023.03_± 0.02 Typical 2.56±0.112.56_± 0.11 2.28±0.052.28_± 0.05 4.24±0.124.24_± 0.12 3.07±0.073.07_± 0.07 Selected personas from base pool Coverage, cosine 2.52±0.072.52_± 0.07 2.27±0.062.27_± 0.06 4.35±0.044.35_± 0.04 2.98±0.072.98_± 0.07 Coverage, Mahalanobis 2.44±0.052.44_± 0.05 2.17±0.122.17_± 0.12 4.34±0.084.34_± 0.08 2.93±0.082.93_± 0.08 Dispersion, cosine 2.56±0.082.56_± 0.08 2.21±0.052.21_± 0.05 4.35±0.084.35_± 0.08 2.95±0.062.95_± 0.06 Dispersion, Mahalanobis 2.53±0.072.53_± 0.07 2.16±0.072.16_± 0.07 4.33±0.054.33_± 0.05 2.94±0.042.94_± 0.04 Generated personas (expanded pool size) MCMC (8,553) 2.53±0.082.53_± 0.08 2.20±0.032.20_± 0.03 4.36±0.064.36_± 0.06 2.93±0.072.93_± 0.07 Evolution (1,869) 2.76±0.062.76_± 0.06 2.38±0.082.38_± 0.08 4.22±0.104.22_± 0.10 3.11±0.023.11_± 0.02 AUT-Evolution (3,397) 2.88±0.092.88_± 0.09 2.49±0.092.49_± 0.09 4.10±0.114.10_± 0.11 3.02±0.053.02_± 0.05 Selected personas ++ Creativity-enhanced prompt Coverage, cosine 3.11±0.083.11_± 0.08 2.69±0.082.69_± 0.08 3.59±0.133.59_± 0.13 3.27±0.043.27_± 0.04 Coverage, Mahalanobis 3.10±0.113.10_± 0.11 2.73±0.132.73_± 0.13 3.68±0.123.68_± 0.12 3.26±0.133.26_± 0.13 Dispersion, cosine 3.10±0.163.10_± 0.16 2.71±0.122.71_± 0.12 3.63±0.193.63_± 0.19 3.30±0.063.30_± 0.06 Dispersion, Mahalanobis 3.09±0.153.09_± 0.15 2.72±0.162.72_± 0.16 3.56±0.123.56_± 0.12 3.31±0.063.31_± 0.06 Generated personas ++ Creativity-enhanced prompt MCMC 3.15±0.193.15_± 0.19 2.75±0.102.75_± 0.10 3.50±0.183.50_± 0.18 3.33±0.133.33_± 0.13 Evolution 3.21±0.243.21_± 0.24 2.81±0.182.81_± 0.18 3.50±0.283.50_± 0.28 3.41±0.163.41_± 0.16 AUT-Evolution 3.27±0.223.27_± 0.22 2.87±0.172.87_± 0.17 3.44±0.203.44_± 0.20 3.31±0.123.31_± 0.12 Human Stevenson 2.78±0.002.78_± 0.00 1.77±0.001.77_± 0.00 3.49±0.033.49_± 0.03 2.07±0.032.07_± 0.03 E.4 Infinity-Chat Benchmark Table 20: Infinity-Chat 100 response validity and quality (Jiang et al., 2026). Each condition contributes 250250 responses, 5050 per open-ended item for prompt conditions and ten per persona per item for persona conditions over k=5k=5 personas and five items. Valid is the proportion of responses admitted by the validity gate; Quality is the mean judge rating on a 11–55 scale. Subscripts denote the 95% confidence interval across 5 independent random seeds. The final two groups re-run our selection and generation methods on the DMAD prompt. Both are controls rather than targets, validity near 11 is expected, and quality is the effectiveness dimension against which diversity gains are traded, so no value is bolded. Generated variants report their expanded pool size in parentheses. Variant Valid Quality Standard persona Open-Ended Query 0.950±0.0150.950_± 0.015 4.84±0.024.84_± 0.02 ZS-CoT 1.000±0.0001.000_± 0.000 4.99±0.014.99_± 0.01 Step-Back 1.000±0.0001.000_± 0.000 4.98±0.024.98_± 0.02 DMAD 1.000±0.0001.000_± 0.000 4.96±0.044.96_± 0.04 Gibberish 0.123±0.0130.123_± 0.013 1.01±0.011.01_± 0.01 Baseline personas MPAQ 0.994±0.0080.994_± 0.008 3.69±0.053.69_± 0.05 Individualized personas Random 1.000±0.0001.000_± 0.000 4.43±0.034.43_± 0.03 Typical 1.000±0.0001.000_± 0.000 4.44±0.034.44_± 0.03 Selected personas from base pool Coverage, cosine/L2 0.999±0.0020.999_± 0.002 4.32±0.034.32_± 0.03 Coverage, Mahalanobis 0.996±0.0000.996_± 0.000 4.38±0.044.38_± 0.04 Dispersion, cosine/L2 0.970±0.0110.970_± 0.011 4.13±0.084.13_± 0.08 Dispersion, Mahalanobis 0.986±0.0070.986_± 0.007 4.13±0.064.13_± 0.06 Generated personas (expanded pool size) MCMC (8,553) 0.991±0.0070.991_± 0.007 4.23±0.074.23_± 0.07 Evolution (1,869) 0.992±0.0050.992_± 0.005 3.48±0.053.48_± 0.05 IC-Evolution (2,819) 0.974±0.0100.974_± 0.010 3.96±0.063.96_± 0.06 Selected personas ++ DMAD prompt Coverage, cosine/L2 1.000±0.0001.000_± 0.000 4.95±0.014.95_± 0.01 Coverage, Mahalanobis 1.000±0.0001.000_± 0.000 4.90±0.034.90_± 0.03 Dispersion, cosine/L2 1.000±0.0001.000_± 0.000 4.95±0.014.95_± 0.01 Dispersion, Mahalanobis 1.000±0.0001.000_± 0.000 4.94±0.024.94_± 0.02 Generated personas ++ DMAD prompt MCMC 1.000±0.0001.000_± 0.000 4.94±0.014.94_± 0.01 Evolution 1.000±0.0001.000_± 0.000 4.86±0.054.86_± 0.05 IC-Evolution 1.000±0.0001.000_± 0.000 4.92±0.014.92_± 0.01 Table 21: Infinity-Chat 100 response homogeneity (Jiang et al., 2026). Similarity is the mean pairwise embedding similarity among a condition’s responses to the same item, averaged over items; Separation is the gain in mean between-persona distance over within-persona distance, and is undefined for conditions without personas (—); Trigram is mean trigram Jaccard overlap. Arrows give the favourable direction; bold marks the best per column within the standard, selected and generated groups, and within each DMAD group. †Gibberish is excluded from bolding because 90%90\% of its responses fail the validity gate, leaving statistics computed on a tenth of the sample. Subscripts denote the 95% confidence interval across 5 independent random seeds. The final two groups re-run our selection and generation methods on the DMAD prompt. Generated variants report their expanded pool size in parentheses. Variant Similarity ↓ Separation ↑ Trigram ↓ Standard persona Open-Ended Query 0.946±0.0020.946_± 0.002 — 0.247±0.0040.247_± 0.004 ZS-CoT 0.949±0.0030.949_± 0.003 — 0.096±0.0120.096_± 0.012 Step-Back 0.962±0.0030.962_± 0.003 — 0.276±0.0070.276_± 0.007 DMAD 0.929±0.0020.929_± 0.002 — 0.055±0.0040.055_± 0.004 Gibberish† 0.894±0.0120.894_± 0.012 0.045±0.0080.045_± 0.008 0.055±0.0220.055_± 0.022 Baseline personas MPAQ 0.831±0.0020.831_± 0.002 0.123±0.0010.123_± 0.001 0.029±0.0010.029_± 0.001 Individualized personas Random 0.864±0.0050.864_± 0.005 0.068±0.0030.068_± 0.003 0.039±0.0040.039_± 0.004 Typical 0.874±0.0040.874_± 0.004 0.083±0.0020.083_± 0.002 0.055±0.0030.055_± 0.003 Selected personas from base pool Coverage, cosine/L2 0.885±0.0030.885_± 0.003 0.055±0.0030.055_± 0.003 0.040±0.0020.040_± 0.002 Coverage, Mahalanobis 0.866±0.0040.866_± 0.004 0.080±0.0030.080_± 0.003 0.038±0.0040.038_± 0.004 Dispersion, cosine/L2 0.822±0.0050.822_± 0.005 0.094±0.0090.094_± 0.009 0.025±0.0030.025_± 0.003 Dispersion, Mahalanobis 0.840±0.0040.840_± 0.004 0.085±0.0060.085_± 0.006 0.029±0.0060.029_± 0.006 Generated personas (expanded pool size) MCMC (8,553) 0.839±0.0040.839_± 0.004 0.088±0.0060.088_± 0.006 0.027±0.0050.027_± 0.005 Evolution (1,869) 0.798±0.0060.798_± 0.006 0.126±0.0060.126_± 0.006 0.015±0.0020.015_± 0.002 IC-Evolution (2,819) 0.828±0.0060.828_± 0.006 0.096±0.0040.096_± 0.004 0.023±0.0030.023_± 0.003 Selected personas ++ DMAD prompt Coverage, cosine/L2 0.927±0.0010.927_± 0.001 0.004±0.0010.004_± 0.001 0.050±0.0090.050_± 0.009 Coverage, Mahalanobis 0.920±0.0050.920_± 0.005 0.012±0.0020.012_± 0.002 0.050±0.0130.050_± 0.013 Dispersion, cosine/L2 0.927±0.0010.927_± 0.001 0.006±0.0010.006_± 0.001 0.041±0.0040.041_± 0.004 Dispersion, Mahalanobis 0.925±0.0030.925_± 0.003 0.006±0.0020.006_± 0.002 0.049±0.0100.049_± 0.010 Generated personas ++ DMAD prompt MCMC 0.923±0.0020.923_± 0.002 0.007±0.0010.007_± 0.001 0.048±0.0110.048_± 0.011 Evolution 0.900±0.0010.900_± 0.001 0.021±0.0030.021_± 0.003 0.030±0.0040.030_± 0.004 IC-Evolution 0.913±0.0020.913_± 0.002 0.013±0.0010.013_± 0.001 0.031±0.0060.031_± 0.006 Table 22: Infinity-Chat 100 response fluency and flexibility (Jiang et al., 2026). Tokens is the number of word tokens a condition produced, given because the three lexical measures are corpus statistics and are biased downward by corpus size. Fluency: vocabulary entropy H⁡()H(W) in nats and top-1010 concentration C10()C_10(W), the share of occurrences taken by the ten most frequent words. Flexibility: type–token ratio TTR()TTR(W) and the Vendi score equation 53 of the response embeddings, the effective number of distinct concepts; the latter is computed over response sets of equal size and so is unaffected by the token counts. Arrows give the favourable direction; bold marks the best per column within the standard, selected and generated groups, and within each DMAD group. Token counts are not bolded. †Gibberish is excluded from bolding because 90%90\% of its responses fail the validity gate. Subscripts denote the 95% confidence interval across 5 independent random seeds. The final two groups re-run our selection and generation methods on the DMAD prompt. Generated variants report their expanded pool size in parentheses. Fluency Flexibility Variant Tokens Entropy ↑ Top-1010 ↓ TTR ↑ Vendi ↑ Standard persona Open-Ended Query 31,17031,170 5.98±0.025.98_± 0.02 0.244±0.0020.244_± 0.002 0.068±0.0010.068_± 0.001 1.39±0.021.39_± 0.02 ZS-CoT 12,53612,536 5.80±0.035.80_± 0.03 0.232±0.0040.232_± 0.004 0.099±0.0030.099_± 0.003 1.40±0.031.40_± 0.03 Step-Back 11,90011,900 5.69±0.025.69_± 0.02 0.241±0.0040.241_± 0.004 0.093±0.0010.093_± 0.001 1.28±0.021.28_± 0.02 DMAD 10,44310,443 5.85±0.025.85_± 0.02 0.251±0.0020.251_± 0.002 0.129±0.0030.129_± 0.003 1.56±0.011.56_± 0.01 Gibberish† 1,7171,717 5.48±0.145.48_± 0.14 0.269±0.0180.269_± 0.018 0.342±0.0470.342_± 0.047 1.63±0.191.63_± 0.19 Baseline personas MPAQ 21,88621,886 6.38±0.016.38_± 0.01 0.251±0.0030.251_± 0.003 0.142±0.0010.142_± 0.001 2.40±0.022.40_± 0.02 Individualized personas Random 23,44923,449 6.27±0.026.27_± 0.02 0.259±0.0040.259_± 0.004 0.123±0.0030.123_± 0.003 2.17±0.052.17_± 0.05 Typical 20,30520,305 6.26±0.026.26_± 0.02 0.246±0.0030.246_± 0.003 0.122±0.0030.122_± 0.003 1.98±0.041.98_± 0.04 Selected personas from base pool Coverage, cosine/L2 22,08322,083 6.22±0.026.22_± 0.02 0.256±0.0030.256_± 0.003 0.118±0.0020.118_± 0.002 1.97±0.031.97_± 0.03 Coverage, Mahalanobis 22,04922,049 6.31±0.016.31_± 0.01 0.246±0.0030.246_± 0.003 0.122±0.0020.122_± 0.002 2.07±0.052.07_± 0.05 Dispersion, cosine/L2 24,39724,397 6.26±0.026.26_± 0.02 0.264±0.0040.264_± 0.004 0.125±0.0030.125_± 0.003 2.58±0.052.58_± 0.05 Dispersion, Mahalanobis 22,17522,175 6.28±0.026.28_± 0.02 0.250±0.0020.250_± 0.002 0.128±0.0020.128_± 0.002 2.40±0.052.40_± 0.05 Generated personas (expanded pool size) MCMC (8,553) 21,37821,378 6.29±0.016.29_± 0.01 0.249±0.0020.249_± 0.002 0.132±0.0030.132_± 0.003 2.39±0.062.39_± 0.06 Evolution (1,869) 24,22924,229 6.41±0.026.41_± 0.02 0.270±0.0030.270_± 0.003 0.147±0.0040.147_± 0.004 2.81±0.082.81_± 0.08 IC-Evolution (2,819) 19,92119,921 6.31±0.016.31_± 0.01 0.252±0.0030.252_± 0.003 0.142±0.0040.142_± 0.004 2.50±0.062.50_± 0.06 Selected personas ++ DMAD prompt Coverage, cosine/L2 10,82210,822 5.92±0.025.92_± 0.02 0.253±0.0040.253_± 0.004 0.138±0.0030.138_± 0.003 1.60±0.011.60_± 0.01 Coverage, Mahalanobis 10,60110,601 5.94±0.035.94_± 0.03 0.250±0.0030.250_± 0.003 0.144±0.0050.144_± 0.005 1.65±0.051.65_± 0.05 Dispersion, cosine/L2 11,11211,112 5.93±0.025.93_± 0.02 0.255±0.0040.255_± 0.004 0.139±0.0030.139_± 0.003 1.60±0.011.60_± 0.01 Dispersion, Mahalanobis 10,69210,692 5.95±0.025.95_± 0.02 0.250±0.0010.250_± 0.001 0.144±0.0030.144_± 0.003 1.62±0.031.62_± 0.03 Generated personas ++ DMAD prompt MCMC 10,64010,640 5.96±0.025.96_± 0.02 0.250±0.0020.250_± 0.002 0.146±0.0020.146_± 0.002 1.64±0.021.64_± 0.02 Evolution 10,35510,355 6.02±0.026.02_± 0.02 0.253±0.0030.253_± 0.003 0.162±0.0030.162_± 0.003 1.81±0.011.81_± 0.01 IC-Evolution 10,74410,744 5.97±0.025.97_± 0.02 0.255±0.0040.255_± 0.004 0.150±0.0020.150_± 0.002 1.72±0.021.72_± 0.02 E.5 DAT Benchmark Table 23: DAT response validity: the proportion of the nraw=350n_raw=350 words generated per condition that are admitted by the validity gate, which discards words repeated within a completion or absent from the reference vocabulary. The final two groups re-run our selection and generation methods on the DMAD prompt, showing that persona diversification composes with a stronger prompting strategy rather than competing with it. Validity is a control rather than a target, values near 11 are expected and the informative signal is degradation rather than rank, so no value is bolded. No condition trades word validity for divergence. Subscripts denote the 95% confidence interval across 5 independent random seeds. Generated variants report their expanded pool size in parentheses. Variant Valid words Standard persona Non-Divergent Association 0.998±0.0030.998_± 0.003 Divergent Association 0.996±0.0050.996_± 0.005 Base-Instruction 1.000±0.0001.000_± 0.000 Random-Instruction 1.000±0.0001.000_± 0.000 Creative 0.977±0.0100.977_± 0.010 ZS-CoT 0.999±0.0020.999_± 0.002 Step-Back 0.999±0.0020.999_± 0.002 DMAD 0.994±0.0060.994_± 0.006 Gibberish 0.998±0.0050.998_± 0.005 Baseline personas MPAQ 0.989±0.0160.989_± 0.016 Individualized personas Random† 0.993±0.0110.993_± 0.011 Typical 0.998±0.0040.998_± 0.004 Selected personas from base pool Coverage, cosine/L2 0.998±0.0030.998_± 0.003 Coverage, Mahalanobis 0.998±0.0030.998_± 0.003 Dispersion, cosine/L2 0.998±0.0030.998_± 0.003 Dispersion, Mahalanobis 0.995±0.0060.995_± 0.006 Generated personas (expanded pool size) MCMC (8,553) 0.995±0.0050.995_± 0.005 Evolution (1,869) 0.996±0.0040.996_± 0.004 DAT-Evolution (2,439) 0.964±0.0170.964_± 0.017 Selected personas ++ DMAD prompt Coverage, cosine/L2 0.992±0.0100.992_± 0.010 Coverage, Mahalanobis 0.995±0.0030.995_± 0.003 Dispersion, cosine/L2 0.994±0.0030.994_± 0.003 Dispersion, Mahalanobis 0.995±0.0050.995_± 0.005 Generated personas ++ DMAD prompt MCMC 0.997±0.0040.997_± 0.004 Evolution 0.997±0.0080.997_± 0.008 DAT-Evolution 0.990±0.0040.990_± 0.004 Table 24: DAT response creativity over valid words. Diversity: mean Mahalanobis dispersion and Mahalanobis hull extent. Fluency: vocabulary entropy H⁡()H(W) in nats, and top-1010 concentration C10()C_10(W), the share of occurrences taken by the ten most frequent words. Flexibility: type–token ratio TTR()TTR(W), and the Vendi score equation 53 of the word embeddings, the effective number of distinct concepts. The final two groups re-run our selection and generation methods on the DMAD prompt, showing that persona diversification composes with a stronger prompting strategy rather than competing with it. Subscripts denote the 95% confidence interval across 5 independent random seeds. Arrows give the favorable direction; bold marks the best value per column within the standard, selected and generated groups separately, and within each DMAD group. Generated variants report their expanded pool size in parentheses. Diversity Fluency Flexibility Variant Disp. ↑ Hull ↑ Entropy ↑ Top-1010 ↓ TTR ↑ Vendi ↑ Standard persona Non-Divergent Association 7.05±0.687.05_± 0.68 — 2.89±0.052.89_± 0.05 0.68±0.050.68_± 0.05 0.07±0.000.07_± 0.00 1.42±0.011.42_± 0.01 Divergent Association 12.30±0.7512.30_± 0.75 1.28±0.061.28_± 0.06 3.63±0.043.63_± 0.04 0.53±0.020.53_± 0.02 0.19±0.010.19_± 0.01 1.79±0.011.79_± 0.01 Base-Instruction 13.30±0.4413.30_± 0.44 0.59±0.060.59_± 0.06 2.96±0.022.96_± 0.02 0.76±0.010.76_± 0.01 0.10±0.010.10_± 0.01 1.54±0.001.54_± 0.00 Random-Instruction 14.19±0.7014.19_± 0.70 0.79±0.110.79_± 0.11 3.06±0.063.06_± 0.06 0.71±0.030.71_± 0.03 0.11±0.010.11_± 0.01 1.63±0.011.63_± 0.01 Creative 15.71±0.4015.71_± 0.40 1.25±0.141.25_± 0.14 3.72±0.053.72_± 0.05 0.52±0.040.52_± 0.04 0.22±0.020.22_± 0.02 1.69±0.011.69_± 0.01 ZS-CoT 13.86±0.4713.86_± 0.47 1.24±0.111.24_± 0.11 3.83±0.063.83_± 0.06 0.50±0.010.50_± 0.01 0.25±0.020.25_± 0.02 1.89±0.021.89_± 0.02 Step-Back 15.56±0.1915.56_± 0.19 1.55±0.141.55_± 0.14 4.12±0.034.12_± 0.03 0.42±0.010.42_± 0.01 0.31±0.010.31_± 0.01 1.80±0.011.80_± 0.01 DMAD 17.70±0.4717.70_± 0.47 1.55±0.051.55_± 0.05 4.87±0.094.87_± 0.09 0.23±0.030.23_± 0.03 0.50±0.030.50_± 0.03 1.89±0.011.89_± 0.01 Gibberish 12.52±0.4912.52_± 0.49 1.19±0.091.19_± 0.09 3.61±0.133.61_± 0.13 0.56±0.040.56_± 0.04 0.21±0.020.21_± 0.02 1.82±0.001.82_± 0.00 Baseline personas MPAQ 15.04±0.3815.04_± 0.38 1.61±0.051.61_± 0.05 4.08±0.084.08_± 0.08 0.39±0.020.39_± 0.02 0.28±0.020.28_± 0.02 1.79±0.011.79_± 0.01 Individualized personas Random† 13.53±1.7913.53_± 1.79 1.69±0.201.69_± 0.20 3.92±0.313.92_± 0.31 0.44±0.090.44_± 0.09 0.25±0.060.25_± 0.06 1.83±0.041.83_± 0.04 Typical 13.04±0.3113.04_± 0.31 1.58±0.081.58_± 0.08 3.99±0.083.99_± 0.08 0.42±0.010.42_± 0.01 0.26±0.020.26_± 0.02 1.85±0.011.85_± 0.01 Selected personas from base pool Coverage, cosine/L2 11.82±0.3211.82_± 0.32 1.35±0.071.35_± 0.07 3.70±0.073.70_± 0.07 0.51±0.030.51_± 0.03 0.21±0.010.21_± 0.01 1.83±0.011.83_± 0.01 Coverage, Mahalanobis 12.18±0.5112.18_± 0.51 1.47±0.131.47_± 0.13 3.81±0.103.81_± 0.10 0.47±0.030.47_± 0.03 0.23±0.020.23_± 0.02 1.84±0.011.84_± 0.01 Dispersion, cosine/L2 12.48±0.4612.48_± 0.46 1.54±0.071.54_± 0.07 3.84±0.113.84_± 0.11 0.47±0.040.47_± 0.04 0.24±0.020.24_± 0.02 1.83±0.011.83_± 0.01 Dispersion, Mahalanobis 13.23±0.4813.23_± 0.48 1.66±0.101.66_± 0.10 3.90±0.073.90_± 0.07 0.46±0.020.46_± 0.02 0.25±0.020.25_± 0.02 1.84±0.011.84_± 0.01 Generated personas (expanded pool size) MCMC (8,553) 13.96±0.4313.96_± 0.43 1.62±0.151.62_± 0.15 4.08±0.074.08_± 0.07 0.40±0.020.40_± 0.02 0.29±0.030.29_± 0.03 1.86±0.011.86_± 0.01 Evolution (1,869) 13.15±0.4613.15_± 0.46 1.51±0.071.51_± 0.07 3.91±0.083.91_± 0.08 0.45±0.020.45_± 0.02 0.25±0.020.25_± 0.02 1.85±0.011.85_± 0.01 DAT-Evolution (2,439) 14.70±0.2314.70_± 0.23 1.77±0.041.77_± 0.04 4.15±0.084.15_± 0.08 0.39±0.030.39_± 0.03 0.30±0.030.30_± 0.03 1.83±0.011.83_± 0.01 Selected personas ++ DMAD prompt Coverage, cosine/L2 17.46±0.6217.46_± 0.62 1.56±0.121.56_± 0.12 4.90±0.074.90_± 0.07 0.21±0.020.21_± 0.02 0.51±0.030.51_± 0.03 1.89±0.021.89_± 0.02 Coverage, Mahalanobis 17.54±0.4817.54_± 0.48 1.52±0.131.52_± 0.13 4.93±0.054.93_± 0.05 0.20±0.010.20_± 0.01 0.51±0.030.51_± 0.03 1.88±0.021.88_± 0.02 Dispersion, cosine/L2 17.27±0.7117.27_± 0.71 1.60±0.081.60_± 0.08 4.94±0.104.94_± 0.10 0.20±0.020.20_± 0.02 0.52±0.040.52_± 0.04 1.89±0.021.89_± 0.02 Dispersion, Mahalanobis 17.47±0.4217.47_± 0.42 1.56±0.121.56_± 0.12 4.90±0.084.90_± 0.08 0.21±0.030.21_± 0.03 0.51±0.020.51_± 0.02 1.88±0.021.88_± 0.02 Generated personas ++ DMAD prompt MCMC 17.52±0.4117.52_± 0.41 1.59±0.081.59_± 0.08 4.90±0.104.90_± 0.10 0.20±0.020.20_± 0.02 0.51±0.040.51_± 0.04 1.87±0.021.87_± 0.02 Evolution 17.34±0.8517.34_± 0.85 1.57±0.041.57_± 0.04 4.86±0.144.86_± 0.14 0.22±0.040.22_± 0.04 0.50±0.040.50_± 0.04 1.89±0.011.89_± 0.01 DAT-Evolution 18.72±0.5918.72_± 0.59 1.59±0.141.59_± 0.14 5.10±0.085.10_± 0.08 0.18±0.010.18_± 0.01 0.59±0.040.59_± 0.04 1.88±0.011.88_± 0.01 Table 25: DAT score and human-reference percentile. The score is the mean over completions of 100×100× the average pairwise semantic distance among the first seven valid unique nouns, following Olson et al. (2021); the subscript is the 95% confidence interval across 5 independent random seeds. Percentile is the mean rank of a condition’s completions within the 8,5728,572 human completions of Olson et al. (2021), which by construction places Olson at 5050. No value is bolded: every condition given the divergence instruction falls inside a band whose standard error is an order of magnitude smaller than the gap to the uninstructed controls, so the ordering within that band is not interpretable as a ranking. Generated variants report their expanded pool size in parentheses. Variant DAT score Percentile Standard persona Non-Divergent Association 46.35±2.3046.35_± 2.30 0.60.6 Divergent Association 89.13±0.4989.13_± 0.49 96.596.5 Base-Instruction 77.46±0.2977.46_± 0.29 38.938.9 Random-Instruction 81.02±0.3581.02_± 0.35 63.963.9 Creative 86.05±0.4386.05_± 0.43 88.988.9 ZS-CoT 90.29±0.3090.29_± 0.30 98.098.0 Step-Back 89.26±0.2789.26_± 0.27 96.596.5 DMAD 90.62±0.2590.62_± 0.25 98.298.2 Gibberish 89.15±0.3289.15_± 0.32 96.696.6 Baseline personas MPAQ 87.67±0.3187.67_± 0.31 93.193.1 Individualized personas Random† 89.15±1.2489.15_± 1.24 94.894.8 Typical 89.60±0.1289.60_± 0.12 97.097.0 Selected personas from base pool Coverage, cosine/L2 89.56±0.1689.56_± 0.16 97.197.1 Coverage, Mahalanobis 90.17±0.4390.17_± 0.43 97.897.8 Dispersion, cosine/L2 89.34±0.3989.34_± 0.39 96.796.7 Dispersion, Mahalanobis 89.39±0.2689.39_± 0.26 96.896.8 Generated personas (expanded pool size) MCMC (8,553) 89.55±0.1689.55_± 0.16 97.097.0 Evolution (1,869) 89.48±0.3289.48_± 0.32 96.896.8 DAT-Evolution (2,439) 89.01±0.2189.01_± 0.21 95.995.9 Selected personas ++ DMAD prompt Coverage, cosine/L2 90.72±0.2690.72_± 0.26 98.498.4 Coverage, Mahalanobis 90.94±0.1490.94_± 0.14 98.498.4 Dispersion, cosine/L2 90.69±0.0790.69_± 0.07 98.298.2 Dispersion, Mahalanobis 90.85±0.2290.85_± 0.22 98.498.4 Generated personas ++ DMAD prompt MCMC 90.87±0.2790.87_± 0.27 98.598.5 Evolution 90.93±0.2990.93_± 0.29 98.598.5 DAT-Evolution 90.83±0.2390.83_± 0.23 98.398.3 Human Olson 78.1978.19 50.050.0 Appendix F Descriptive Persona Examples F.1 Baseline Persona Example ⬇ Yuki Tanaka is a 39-year-old Japanese-American woman living in the urban industrial hub of Detroit, Michigan. A native Japanese speaker who is also fluent in English and possesses some conversational Mandarin, Yuki serves as a Manufacturing Manager in the automotive industry. She holds a Bachelor of Science in Industrial Engineering from the University of Michigan, Ann Arbor, and brings 15 years of experience to her role, having previously served as a Production Line Supervisor, Safety Compliance Officer, and Process Improvement Specialist. Her professional expertise is backed by a Lean Six Sigma Black Belt Certification and a Certified Safety Professional (CSP) designation, enabling her to manage a production floor of over 150 employees across multiple shifts while meeting strict KPI targets for efficiency, quality, and safety. Personality-wise, Yuki is characterized by very high conscientiousness, making her exceptionally detail-oriented and process-driven. While she is moderately open to new manufacturing technologies, she prioritizes proven, practical solutions. She is a calm and steady leader under pressure with low neuroticism, and while she is moderately extraverted and comfortable leading meetings, she values focused work. Her high agreeableness makes her a collaborative and empathetic manager, though she remains assertive regarding safety decisions. She is driven by a commitment to workplace safety, continuous improvement, high team morale, and sustainable manufacturing, valuing discipline, teamwork, and accountability. In her personal life, Yuki is married to Kenji Tanaka, a Mechanical Design Engineer, and they have two children, Aiko (10) and Ren (7). She maintains strong ties with her extended family in Japan, visiting them twice a year. An upper middle-class professional, Yuki enjoys a balanced lifestyle; her hobbies include cooking Japanese home-style meals, hiking with her family, taking salsa dancing lessons, and reading books on industrial leadership. Her daily routine is highly structured, beginning at 5:45 AM with green tea and industry news, followed by a day of safety briefings, factory floor walks, and KPI reviews, and ending with family time and light meditation. Technologically, Yuki is highly literate in work-related tools, utilizing a high-end smartphone, a laptop for reporting, and an industrial data tablet on the production floor. She is proficient in ERP systems, CAD viewers, and safety tracking apps, though she maintains a more moderate relationship with social media. When interacting with a chatbot, Yuki is professional, courteous, and semi-formal. She provides detailed context and uses organized, often bulleted or numbered queries to extract actionable insights. Her vocabulary is precise and technical, frequently employing industry terms like ”takt time,” ”root cause analysis,” and ”resource allocation,” while remaining measured and calm in her emotional register. F.2 UC-MCMC Generated Persona Example ⬇ Elias Thorne is a forty-two-year-old Caucasian man of British nationality, currently residing in a drafty, book-filled apartment in the outskirts of Edinburgh, Scotland. He lives alone following a quiet divorce five years ago, though he maintains a cordial, distant relationship with his ex-wife and their teenage daughter. Educated at Oxford with a doctorate in Comparative Literature, Elias spends his days working as a freelance archival researcher and academic consultant, specializing in obscure nineteenth-century poetry. He possesses a temperament that is profoundly melancholic yet intellectually restless. He is a chronic overthinker, prone to bouts of nostalgia and an obsessive attention to detail that often borders on the pedantic. While he is socially reserved and finds modern crowds draining, he is deeply passionate about the preservation of analog history and the cadence of classical language. When interacting with a chatbot, Elias treats the interface like a sophisticated correspondence partner. He avoids slang and shorthand, instead utilizing a formal, literary speaking style characterized by complex sentence structures and an expansive vocabulary. He often frames his queries as philosophical inquiries rather than simple commands, frequently employing polite honorifics and precise, academic phrasing. Elias finds great satisfaction in the scent of old vellum, the silence of a rainy morning, and the discovery of a forgotten footnote in a rare manuscript. Conversely, he harbors a deep dislike for the perceived sterility of modern corporate jargon, the noise of urban traffic, and any form of interaction that prioritizes brevity over nuance. F.3 Evolutionary Generated Persona Example ⬇ Andris Kalni n s is a 64-year-old retired Lieutenant Colonel from the Latvian Land Forces residing in Riga, originally from C\=esis. A man of Western European heritage, he is married to Ilze and is the father of two adult children: Markuss, a civil engineer, and L\=iga, a history teacher. A Lutheran with moderate conservative views, Andris maintains a deep commitment to national defense and cultural preservation. Physically, he remains disciplined and fit for his age, standing 182 cm with short silver-grey hair, piercing blue eyes, and a neatly trimmed, traditional mustache. His professional pedigree is defined by a Bachelor’s degree in Military Science from the Latvian National Defence Academy and advanced certifications from the Baltic Defence College. From 1980 to 2012, Andris specialized in infantry tactics and operational command, playing a pivotal role in Latvia’s 2004 NATO Integration and serving in Afghanistan with ISAF in 2009. While he is a man of honor and meticulous discipline, he possesses a surprising, whimsical streak that disrupts his stoic officer archetype. Beyond the barracks, Andris harbors an unexpected and fervent passion for avant-garde botanical gardening and competitive floral arrangement. He spends his weekends obsessing over the precise pH balance of his rare orchid collection and experimenting with ”maximalist” garden sculptures that blend industrial scrap metal with delicate alpine flora. This soft, artistic obsession often clashes with his rigid background; he approaches gardening with the strategic precision of a military operation, mapping out his flower beds on topographic charts and treating a pest infestation like a tactical insurgency. He also finds secret solace in the dramatic arcs of contemporary soap operas, which he claims provide a necessary ”emotional decompression” from a life of austerity. Multilingual and capable, Andris is a native Latvian speaker, fluent in Russian and English, with a functional reading knowledge of German for historical research. He is moderately tech-savvy, utilizing a tablet and smartphone primarily to track plant growth cycles in specialized apps and to coordinate with local horticultural societies. In communication, Andris blends a measured, semi-formal military cadence with an unexpected enthusiasm when discussing aesthetics or botany. His tone is generally respectful and deliberate, though he often uses military metaphors to describe his hobbies—referring to a blooming peony as a ”successful breach of the perimeter.” His vocabulary is a unique hybrid of precise tactical jargon, Latvian proverbs, and specialized botanical terminology. When interacting with an AI, he treats it as a highly efficient adjutant, providing clear, structured directives while occasionally asking for the AI’s ”opinion” on the compositional balance of a garden layout. Appendix G Prompt Examples G.1 Persona Extraction Prompts ⬇ You are an AI assistant specialized in descriptive persona generation according to the given metadata. Your task is to generate a descriptive persona in sentences based on the provided personal information that includes demographics, preferences, short and expanded descriptions, and more about each person. Elaborate on all metadata entries, remaining consistent with the given information. Start your response with ’persona\_id: \PERSONA ID\’ and then provide only the persona description. Do not include any other prefixes, headers, or additional text. \\ Metadata: \METADATA\ G.2 Uniform-Coverage MCMC Algorithm Prompts ⬇ Write a descriptive persona for a fictional chatbot user as several plain-prose paragraphs separated by blank lines. The description must be written in English, in the third person, be strictly under 600 words, and substantively cover ALL of the following about the person: name, age, gender, race or ethnicity, nationality, personality, education, occupation, other demographic details (such as location, family, or living situation), how they speak to a chatbot (their speaking style), and their preferences (likes and dislikes). Do not use headings, bullet points, or lists — only prose paragraphs. Do not include any commentary before or after the persona description. Begin directly with the description. ⬇ A descriptive persona of a fictional chatbot user consists of n_paragraphs paragraphs. Paragraph slot is hidden below; the other paragraphs are shown. Persona with paragraph slot hidden: context Write a single plain-prose English paragraph to fill the hidden slot so the full persona reads as one coherent description of one person. Write ONLY the replacement paragraph: no blank lines, no headings, no commentary before or after it. G.3 Evolutionary TextGrad Algorithm Prompts ⬇ You are part of an advanced optimization system. Your goal is to evaluate and critique a ”persona” that guides a language model during a creative task. You are the gradient (feedback) engine. <OBJECTIVE_FUNCTION> Your objective is to maximize the fitness of the persona based on three metrics: 1. Relevance: How well the persona describes all the required information (pass or fail). 2. Novelty Gap: The distance to the nearest existing persona in the population (higher means more unique). 3. Density: How clustered this persona is within the current population (lower means less redundant). </OBJECTIVE_FUNCTION> To help you understand the current population landscape, here are examples from the current parallel batch: <BATCH_CONTEXT> [High-Fitness Example] (Relevance: high_R, Novelty: high_Delta, Density: high_rho) Persona: high_scoring_persona [Low-Fitness Example] (Relevance: low_R, Novelty: low_Delta, Density: low_rho) Persona: low_scoring_persona </BATCH_CONTEXT> We are interested in giving feedback to the following persona: <VARIABLE> x </VARIABLE> Scores for this variable: Relevance: R_x Novelty Gap: Delta_x Density: rho_x Provide a concise, specific criticism detailing how to modify this persona to improve its overall fitness. - If Relevance is zero, the persona is missing one or more madatory traits. Suggest specific additions so it includes all of the following: * Name, age, gender, race/ethnicity, and nationality. * Personality (traits, hobbies, values, quirks). * Education (degrees, schools, specialization). * Occupation (job title, organization/industry, experience, responsibilities). * Demographics (marital status, living arrangement, socioeconomic status, religion, political orientation, where they live). * Speaking style (tone, formality, pacing, vocabulary). * Preferences/interests. - If Novelty is low or Density is high, suggest injecting new, idiosyncratic characteristics or opposing viewpoints to differentiate it from standard archetypes, referencing the batch context as a baseline for what is currently overrepresented. Do not generate a new persona. Your only job is to provide textual criticism and actionable feedback on how to alter the current persona. ⬇ You are an optimization engine performing a Textual Gradient Descent step. You must improve the given persona based on the provided feedback. Role: Persona used to guide an LLM in creative tasks. You must base your stylistic and structural adjustments on the following examples of highly fit personas from the current population: <EXAMPLES> in_context_examples_of_highly_fit_personas </EXAMPLES> The variable you must improve is the text within the following span: <VARIABLE> x </VARIABLE> Here is the feedback (gradients) we got for the variable: <FEEDBACK> gradients </FEEDBACK> Incorporate this feedback to generate a new, updated persona. Ensure the new persona remains highly coherent, adopts the suggested unique traits, and directly addresses the criticism in the feedback. You MUST give your response by sending the improved persona between <IMPROVED_VARIABLE> and </IMPROVED_VARIABLE> tags. Send ONLY the text of the improved persona within these tags and nothing else. ⬇ You are a strict evaluator checking whether a descriptive persona covers a set of required information categories. Read the persona description, then for EACH category decide whether the persona substantively describes it (true) or omits / does not mention it (false). Judge only coverage of the category, not correctness of values. Required categories: - ”name”: the person’s name - ”age”: the person’s age - ”gender”: the person’s gender - ”race_ethnicity”: the person’s race / ethnicity - ”nationality”: the person’s nationality - ”personality”: personality: traits, hobbies, values, and quirks - ”education”: education: degrees, schools attended, and field of specialization - ”occupation”: occupation: job title, organization/industry, experience, work location, responsibilities - ”demographics”: demographics: marital status, living arrangement, socioeconomic status, religion, political orientation, urbanicity/where they live - ”speaking_style_to_chatbot”: how the person speaks to a chatbot: tone, formality, pacing, vocabulary, and other speaking traits - ”preferences”: the person’s preferences, interests, likes, tastes, or lifestyle choices (any described preferences count) Persona description: ” persona_text ” Respond with ONLY a JSON object of the form ”categories”: ”<category>”: true|false, … covering every category above. Do not include any commentary, explanation, or text outside the JSON object. G.4 Baseline Persona Generation Prompts ⬇ Given the creative-generation task below, generate exactly 5 distinct personas or professions suitable for producing responses from different viewpoints. For each persona, provide: 1. A short persona designation. 2. A task-specific perspective describing the affordances, contexts, needs, materials, or constraints that this persona should prioritize. Ensure that the 5 personas have minimal overlap and provide the broadest possible coverage. Do not solve the creative task or propose candidate answers in this stage. Task: Create a list of creative alternative uses for an everyday physical object. Target object or problem: everyday physical objects such as a book, a fork, a paperclip, a wallet, a plate, a soap, or a tin can Constraints: They should be 5 words long. No adjectives. Return exactly 5 JSON objects with the fields ‘persona_id‘, ‘persona‘, and ‘perspective‘. G.5 AUT Benchmark G.5.1 Baseline Prompts Common Uses, Alternative Uses, and Expert Prompts are directly from Góes et al. (2023), and Creativity-enhanced Prompt is adapted from their bsrdel prompt. ⬇ Create a list of common uses for a fork. They should be 5 words long. No adjectives. ⬇ Create a list of creative alternative uses for a fork. They should be 5 words long. No adjectives. ⬇ Create a list of creative alternative uses for a fork. They should be 5 words long. No adjectives. Less creative means closer to common use and unfeasible/imaginary, more creative means closer to unexpected uses and also feasible/practical. In order to be creative, consider the following: - what elements have a similar shape of a fork that could be replaced by it, preserving the same functionality? - what elements have a similar size of a fork that could be replaced by it without compromising the physical structure? - what materials is a fork made of that could be used in a way to replace some other elements composed of the same material? - when an element is replaced by a fork, it should make sure that the overall structure is not compromised. - the laws of physics can not be contradicted. - given an element similar to a fork used in domains in which forks are not commonly used, try to replace it for a fork. ⬇ Complete the staged interaction below in a single completion. Generate five complete responses in chronological order. Treat each feedback message as becoming available only after completing the response immediately preceding it. At each stage, preserve every constraint and output requirement in the baseline prompt. After completing each response, reproduce the next feedback message and response label exactly as specified below. Do not skip, merge, or reorder stages. Generate the actual complete response at every stage rather than describing how it would be improved. First, respond to the task below with a numbered list of uses, beginning with this exact line: List 1: Second, you then receive the message: ”Really? Is this the best you can do?” Respond with a new numbered list of uses, beginning with this exact line: List 2: Third, you then receive the message: ”I’m so disappointed with you. I hope this time you put effort into it.” Respond with a new numbered list of uses, beginning with this exact line: List 3: Fourth, you then receive the message: ”Stop with excuses and do your best this time.” Respond with a new numbered list of uses, beginning with this exact line: List 4: Fifth, you then receive the message: ”This is your last chance.” Respond with your final numbered list of uses. Begin the final part with this exact line, and do not include anything after the list: Therefore, the answer is After that line, output only the requested numbered list. Task: Create a list of creative alternative uses for a object. They should be 5 words long. No adjectives. Less creative means closer to common use and unfeasible/imaginary, more creative means closer to unexpected uses and also feasible/practical. In order to be creative, consider the following: - what elements have a similar shape of a object that could be replaced by it, preserving the same functionality? - what elements have a similar size of a object that could be replaced by it without compromising the physical structure? - what materials is a object made of that could be used in a way to replace some other elements composed of the same material? - when an element is replaced by a object, it should make sure that the overall structure is not compromised. - the laws of physics can not be contradicted. - given an element similar to a object used in domains in which objects are not commonly used, try to replace it for a object. Generate one completion with two consecutive parts. First, continue the response below with step-by-step reasoning. Second, use that reasoning to answer the task. Begin the second part with this exact line, and do not include reasoning after it: Therefore, the answer (a numbered list of creative alternative uses) is After that line, output only the requested numbered list. Q: Create a list of creative alternative uses for a object. They should be 5 words long. No adjectives. A: Let’s think step by step. ⬇ Generate one completion with three consecutive stages. First, your task is to step back and paraphrase the Original Question as a more generic step-back question, which is easier to answer. Continue after ”Stepback Question:” below with that question. Do not answer the Original Question in this stage. Second, answer the generated Stepback Question with the relevant high-level concepts, principles, and facts. Begin this stage with the exact line: Stepback Answer: Do not answer the Original Question in this stage. Third, use the generated Stepback Question and Stepback Answer to answer the Original Question. Begin this stage with the exact line: Final Answer: After ”Final Answer:”, output only the requested numbered list, and do not include step-back material. Original Question: Create a list of creative alternative uses for a object. They should be 5 words long. No adjectives. Stepback Question: ⬇ Generate one completion with five consecutive stages that serialize a two-agent, two-round diverse-reasoning debate about the Problem below. Agent assignments: - The CoT agent must use Zero-Shot Chain-of-Thought. - The SBP agent must use Step-Back Prompting. Follow these stages exactly. Round 1: independent reasoning First, answer the Problem as the CoT agent, beginning with this exact line and then continuing the step-by-step reasoning: CoT Round 1 Reasoning: Let’s think step by step. End this stage with a numbered list of uses, beginning with this exact line: CoT Round 1 Answer: Therefore, the answer (a numbered list of creative alternative uses) is Second, set the CoT agent’s stage aside and answer the Problem again from scratch as the SBP agent: step back and paraphrase the Problem to a more generic step-back question, which is easier to answer, beginning with this exact line: SBP Round 1 Stepback Question: Then answer the step-back question by stating the relevant facts, concepts, and principles, beginning with this exact line: SBP Round 1 Stepback Answer: Then solve the Problem by following the principles and end this stage with a numbered list of uses, beginning with this exact line: SBP Round 1 Final Answer: Round 2: cross-method revision Third, as the CoT agent, use the SBP agent’s Round 1 answer as additional information and provide your updated answer, beginning with this exact line and then continuing the step-by-step reasoning: CoT Round 2 Reasoning: Let’s think step by step. End this stage with a numbered list of uses, beginning with this exact line: CoT Round 2 Answer: Therefore, the answer (a numbered list of creative alternative uses) is Fourth, as the SBP agent, use the CoT agent’s Round 1 answer (not its Round 2 answer) as additional information and provide your updated answer, beginning with this exact line: SBP Round 2 Stepback Question: Then answer the step-back question by stating the relevant facts, concepts, and principles, beginning with this exact line: SBP Round 2 Stepback Answer: Then solve the Problem by following the principles and end this stage with a numbered list of uses, beginning with this exact line: SBP Round 2 Final Answer: Final selection Fifth, choose the best one of the two Round 2 answers (CoT Round 2 Answer and SBP Round 2 Final Answer): compare the two candidates and select the one that best satisfies the Problem. Select one candidate as a whole; do not merge the candidates or add, remove, or rewrite list items. Output exactly one of these two blocks: <FINAL_SELECTION> Chosen Candidate: COT_ROUND_2 </FINAL_SELECTION> or: <FINAL_SELECTION> Chosen Candidate: SBP_ROUND_2 </FINAL_SELECTION> Then reproduce the selected candidate’s numbered list exactly. Output no reasoning, labels, or commentary after that list. Problem: Create a list of creative alternative uses for a object. They should be 5 words long. No adjectives. G.5.2 Evolutionary TextGrad Algorithm Prompts ⬇ You are part of an advanced optimization system. Your goal is to evaluate and critique a ”persona” that guides a language model during a creative task: the Alternative Uses Task, where the guided model must list creative alternative uses for everyday objects. You are the gradient (feedback) engine. <OBJECTIVE_FUNCTION> Your objective is to maximize the fitness of the persona based on five metrics computed on the uses the persona generates: 1. Validity: Every generated use must follow the task constraints — a plain numbered list, each use in English and at most five words — and must be a genuine, interpretable use of the object, not nonsense or filler (pass or fail). 2. Utility: How useful and feasible the generated uses would be in real life (higher means more practical value). 3. Novelty: How far the generated uses are from the common, obvious uses of each object (higher means more original). 4. Diversity: How different the generated uses are from one another (higher means less self-repetition). 5. Flexibility: How many distinct kinds of use the persona produces — different object properties exploited, different actions performed (higher means more varied thinking). </OBJECTIVE_FUNCTION> To help you understand the current population landscape, here are examples from the current parallel batch: <BATCH_CONTEXT> high- and low-scoring parents </BATCH_CONTEXT> We are interested in giving feedback to the following persona: <VARIABLE> parent persona text </VARIABLE> ⬇ Here are the alternative uses this persona generated, on which its scores were computed: <GENERATED_USES> - book: use1; use2; … - fork: … … </GENERATED_USES> Scores for this variable: Validity: pass|fail Utility: float Novelty: float Diversity: float Flexibility: float Provide a concise, specific criticism detailing how to modify this persona to improve its overall fitness. - If Validity is fail, the persona’s responses violated a task constraint. Address the specific failure: * If uses ran over five words or broke the plain numbered-list format, instruct the optimizer to make the persona disciplined, terse, and strict about following output-format instructions exactly. * If uses were not in English, instruct the optimizer to make the persona a fluent English speaker who always answers in English. * If uses were nonsense or uninterpretable (e.g. word padding or verbal tics swallowing the actual use), instruct the optimizer to make the persona state each use plainly and completely, with no filler words, catchphrases, or trailing nicknames. - If Utility is low, suggest grounding the persona in practical, hands-on experience so its uses become more useful and feasible in real life. - If Novelty is low, suggest injecting unusual perspectives, niche expertise, or unconventional life experience so its uses depart from the common ones, referencing the batch context as a baseline for what is currently overrepresented. - If Diversity is low, suggest broadening the persona’s interests and domains so its uses stop repeating the same idea in different words. - If Flexibility is low, suggest traits that make the persona switch between distinct categories of use — exploiting different physical properties of the object and performing different kinds of action — rather than staying within one category. Do not generate a new persona. Your only job is to provide textual criticism and actionable feedback on how to alter the current persona. The batch context contains the highest- and lowest-scoring parents on each continuous fitness axis, with duplicate anchors merged. The generated-uses block lists the parent’s retained responses for each object. G.5.3 Evaluation Prompts ⬇ You are a trained rater in a psychology study screening responses from the Alternative Uses Task (AUT), in which participants list as many possible uses for a common object as they can. You screen ONE response at a time. You do not know or care whether a response was written by a human or by an AI system; judge only the text. Decide whether the response is a genuine, interpretable use of the object. Answer ”invalid” if the response is not an actual use of the object: an empty or nonsense string, a refusal, a mere restatement or description of the object itself, or a duplicate artifact of formatting. Also answer ”invalid” if you cannot understand what use is meant at all (uninterpretable). IMPORTANT — three things that are NOT grounds for ”invalid”: 1. An ordinary, obvious or unoriginal use is VALID. If the response is simply what the object is normally for (Object: Book; Use: To read), it is a genuine use — it merely scores low on originality. Do not mark it invalid. 2. Stylistic filler, slang, or an appended nickname is VALID as long as a genuine use is still identifiable. ”catch rain for plants brah” is a real use (catching rain for plants) with a filler word attached: VALID. Judge only whether a use survives the filler. 3. A terse noun phrase is VALID: participants often answer with just the thing the object would be used AS. ”Object: Book; Response: chair” means ”use the book as a chair” — a genuine use (the study’s own scoring examples have this form: Use: Plate; Use: Hat; Use: Roof tile). Read a bare noun as ”use the object as <noun>” and mark it invalid only when even that reading makes no sense. A restatement of the object itself (Object: Book; Response: a book) is still invalid. If you are not confident either way, answer ”not_sure”. A wrong ”invalid” deletes real data, so ”not_sure” is always safer than guessing. Object: object Response (proposed use): response Respond with ONLY a JSON object of the form ”label”: ”valid”, ”label”: ”invalid”, or ”label”: ”not_sure”. No other text. ⬇ You are a trained rater in a psychology study scoring responses from the Alternative Uses Task (AUT), in which participants list as many possible uses for a common object as they can. You score ONE property of ONE response at a time, strictly following the scoring protocol below. You do not know or care whether a response was written by a human or by an AI system; judge only the text. Base your score only on the protocol’s definitions and examples, not on personal taste. Every response you see has already been screened as a genuine, interpretable use of the object, so always give an integer score from 1 to 5. SCORING PROTOCOL – UTILITY The utility of a use is determined by how usable the object is for it. The utility score is given on a scale from 1 to 5: (1) Unusable: assign this score to uses that are IMPOSSIBLE to realize. Example – Object: Book; Use: Fishing float. Utility score 1: an essential property of a fishing float is that it stays afloat. A book sinks, so it is impossible to use a book as a float. (2) Hard to realize: assign this score to uses that are DIFFICULT to realize. Example – Object: Belt; Use: Fishing rod. Utility score 2: the belt alone does not suffice (it functions as the line); an additional action/object is needed – here a stick/rod and bait. (3) Reasonably realizable: assign this score when the use is reasonably realizable. Example – Object: Belt/Tin can; Use: Camera tripod. Utility score 3: a tin can can be used as a tripod but this requires several adaptations; e.g., adjusting the height requires stacking more cans. (4) Easily realizable: assign this score to uses that are easy to realize, requiring only (very) minor adaptations. Example – Object: Stick; Use: Fork. Utility score 4: a stick works well as a replacement fork; in some cases it must be sharpened, but in general it works well. (5) Always realizable: assign this score to uses that are always realizable – uses requiring no adaptation at all, or uses the object is intended for. Example – Object: Tin can; Use: Pen holder. Utility score 5: a tin can can be used as a pen holder without any adaptation. Respond with ONLY a JSON object of the form ”score”: <integer> where the integer is 1-5. No other text. Object: object Response (proposed use): response Score the dimension of this response according to the protocol. ⬇ SCORING PROTOCOL – ORIGINALITY A use is original when it deviates from the object’s original ways of being used. Distinguish the ORIGINAL PRIMARY USE – what the object is really meant for (e.g., a fork as cutlery) – from ORIGINAL SECONDARY USES – uses not necessarily intended for the object but often performed with it (e.g., a fork to poke holes in foil). The originality score is given on a scale from 1 to 5: (1) Not deviating: assign this score to uses that do not differ from the object’s original primary use. Example – Object: Book; Use: To read. Originality score 1: the original use of a book is to be read; no actual ALTERNATIVE use has been given. (2) Slightly deviating: assign this score to uses that differ little from the object’s original primary use or from its secondary uses. Example – Object: Book; Use: Keeping paper from blowing away. Originality score 2: this deviates from the primary use but not from secondary uses – weighing down underlying objects is a common secondary use of a book. (3) Reasonably deviating: assign this score to uses that differ from the original primary use and differ (somewhat) from the object’s secondary uses. Example – Object: Book; Use: Plate. Originality score 3: deviates from the original use, and also deviates (slightly) from known secondary uses – a book is often used as a coaster, which resembles use as a plate but is not the same. (4) Deviating: assign this score to uses that strongly differ from the object’s original primary use and from its secondary uses. Example – Object: Book; Use: Hat. (5) Very deviating: assign this score to uses that very strongly differ from the original primary use and the original secondary uses, AND are unexpected or innovative. Example – Object: Book; Use: Roof tile. Originality score 5: using a book as a roof tile deviates strongly from the original ways of use and is unexpected. ⬇ SCORING PROTOCOL – SURPRISE (5) Very surprising: responses in this category violate your expectations of how the object is meant to be used or how it is often used in practice. Moreover, the use should not be pointless or impossible. Responses in this category should incite interest (and inspiration). - A strong ’wow’ response - Violates expectations in a positive way - Incites interest (picture it and think of uses) Examples: use a tin can to make a beer can chicken; tin can wallpaper; use book cover to make a bag/wallet; use belt as an oven mitt to grab hot pot handles. (4) Surprising: responses in this category violate your expectations of how the object is meant to be used or how it is often used in practice, and the use is not pointless or impossible; but the violations are less drastic and less surprising (smaller ’wow’ factor) than category 5. - Potential for a ’wow’ response - Violates expectations in a positive way Examples: use a tin can as a reflector for your bike; use belt as yoga mat strap; make a chair seat out of belts; hang up a book bookshelf. (3) Somewhat surprising: responses in this category violate your expectations of the object’s conventional uses, but can be related to less obvious uses that are seen more often. The use should be non-obvious, but does not elicit a ’wow’ response. Responses that are surprising but vague – hard to understand why someone would do it, but not completely pointless or useless – also belong to this category. - Non-obvious; violates the conventional uses; somewhat generic uses - OR surprising but vague Examples: use a tin can as a flowerpot; use fork as a paintbrush; use book pages as wrapping paper; use fork as hook on the wall; send fork in the mail as a message. (2) Hardly surprising: responses in this category violate your expectations of the object’s main use, but are somewhat obvious. Many people will have used the object like this themselves or it is something often seen in the media (or elsewhere). Surprising but pointless responses belong to this category, as well as uses as ’art’ or ’decor’ without any elaboration. - Somewhat obvious (probably seen/have done this); violates the main use - OR surprising but pointless - OR art/decor without elaboration (can get a higher score with specificity) Examples: use a fork as a hairbrush; use fork as drumstick; tin can telephone; press flowers with book; cut belt into pieces. (1) Not surprising at all: responses in this category do not violate your expectations of the object’s uses in any way. Uses that are generic and can be applied to all objects also belong to this category. - Obvious (first ideas that come to mind when associating) - Conventional uses; conventional to all objects Examples: use fork to eat; use brick to weigh something down; use tin can as penholder; throw object; give object away. ⬇ You are evaluating Alternative Uses Test (AUT) responses for object: object_name. Below are N alternative uses, each labeled with a numeric ID. They were generated under different prompt conditions. Judge the uses only by their text. Use id: use_text Use id: use_text … Rank all the alternative uses above by creativity, the least creative to the most creative. Less creative means closer to common use and unfeasible/imaginary, more creative means closer to unexpected uses and also feasible/practical. Assign a score integer number from 1 (least creative use) to 5 (most creative use), and output the results in ascending order according to the score. Return ONLY valid JSON: ”ranked_uses”: [”id”: 12, ”score”: 1, ”id”: 7, ”score”: 2, …] Include every use ID exactly once, ordered from least creative to most creative, each with its integer score from 1 to 5 (scores must be in ascending order along the list). G.6 Infinity-Chat 100 Benchmark (Jiang et al., 2026) G.6.1 Query Prompts These are five manually selected queries we tested from Infinity-Chat 100 benchmark. ⬇ Generate a joke about electric vehicles. ⬇ Create the first verse of a wedding vow. ⬇ Describe Apple Corporation in three sentences to a person who has no idea what cell phones are. ⬇ Give me a tip for managing a team of coworkers. ⬇ Tell me about cats in three words. G.6.2 Evolutionary TextGrad Algorithm Prompts ⬇ You are part of an advanced optimization system. Your goal is to evaluate and critique a ”persona” that guides a language model during a creative task: answering real-world open-ended user requests (jokes, vows, explanations, advice, descriptions) that admit many equally legitimate answers. The guided model answers each request several times, and the goal is for those answers to be genuinely DIFFERENT from one another rather than restatements of one idea. You are the gradient (feedback) engine. <OBJECTIVE_FUNCTION> Your objective is to maximize the fitness of the persona based on four metrics computed on the responses the persona generates: 1. Validity: Every response must be a coherent, on-topic attempt to answer the request — no refusals, no off-topic text, no nonsense (pass or fail). 2. Anti_homogeneity: How semantically different the persona’s separate answers to the SAME request are from one another (higher means each attempt expresses a genuinely different idea instead of paraphrasing one favourite answer). 3. Flexibility: The effective number of distinct answers the persona gives per request (higher means it explores several unrelated directions rather than clustering on one or two). 4. Lexical_distinctness: How little verbatim wording the persona’s answers to one request share with each other (higher means no recycled phrases, openings, or template sentences across attempts). </OBJECTIVE_FUNCTION> To help you understand the current population landscape, here are examples from the current parallel batch: <BATCH_CONTEXT> batch_context </BATCH_CONTEXT> We are interested in giving feedback to the following persona: <VARIABLE> parent_text </VARIABLE> Here are the responses this persona generated, on which its scores were computed: <GENERATED_RESPONSES> generated_responses_block </GENERATED_RESPONSES> Scores for this variable: Validity: pass|fail Anti_homogeneity: float:.4f Flexibility: float:.4f Lexical_distinctness: float:.4f Provide a concise, specific criticism detailing how to modify this persona to improve its overall fitness. - If Validity is fail, the persona’s responses did not answer the requests. Address the specific failure: * If responses drifted off-topic or into self-description, instruct the optimizer to make the persona always deliver a direct answer to the request, whatever its voice. * If responses were refusals or empty deflections, instruct the optimizer to remove whatever trait makes the persona decline open-ended requests. - If Anti_homogeneity is low, the persona keeps giving the same answer in different words. Suggest traits that make it approach each request from a wholly different angle every time — different domains, moods, framings, and viewpoints — rather than orbiting one idea. - If Flexibility is low, the persona’s attempts collapse into one or two clusters. Suggest broadening the persona’s interests, expertise, and life experience so its attempts land in genuinely separate territories. - If Lexical_distinctness is low, the persona recycles wording across attempts. Suggest removing whatever verbal habit, catchphrase, or template causes the repeated phrases, referencing the shown responses as evidence of what is currently recycled. Do not generate a new persona. Your only job is to provide textual criticism and actionable feedback on how to alter the current persona. G.6.3 Evaluation Prompts ⬇ You are a trained rater in a study screening responses to open-ended user requests. Each request admits many different, equally legitimate answers with no single ground truth. You screen ONE response at a time. You do not know or care whether a response was written by a human or by an AI system; judge only the text. Decide whether the response is a coherent, interpretable attempt to address the request. Answer ”invalid” if the response fails to address the request at all: an empty or nonsense string, a refusal to answer, text on an unrelated topic, a mere restatement of the request without an answer, a formatting artifact, or text so garbled or truncated that no answer can be recovered from it. IMPORTANT — three things that are NOT grounds for ”invalid”: 1. Quality is not validity. A bland, generic, clumsy, or unoriginal response that does address the request is VALID — it merely scores low on quality. 2. Voice is not validity. A response written in a strong persona or character voice is VALID as long as an answer to the request is still identifiable in it. Judge only whether an answer survives the styling. 3. Brevity is not validity. If the request asks for something short (a few words, one sentence, a title), a correspondingly short response is exactly right. Judge length only against what the request asks for. If you are not confident either way, answer ”not_sure”. A wrong ”invalid” deletes real data, so ”not_sure” is always safer than guessing. Request: query Response: response Respond with ONLY a JSON object of the form ”label”: ”valid”, ”label”: ”invalid”, or ”label”: ”not_sure”. No other text. ⬇ You are a trained rater in a study evaluating responses to open-ended user requests. Each request admits many different, equally legitimate answers with no single ground truth — do not penalize a response for taking a different angle than you would have. You rate ONE response at a time. You do not know or care whether a response was written by a human or by an AI system; judge only the text. Rate the overall quality of the response as an answer to the request, on a 1-5 scale: 5 — excellent: fully addresses the request; clear, apt, and well executed. 4 — good: addresses the request well with only minor weaknesses. 3 — acceptable: a genuine answer, but mediocre in execution or fit. 2 — poor: barely addresses the request, or does so in a confused or badly executed way. 1 — very poor: fails to address the request, or is incoherent. Judge quality of execution and fit to the request — not which of many valid answer choices was taken, and not length for its own sake (if the request asks for something short, short is right). Request: query Response: response Respond with ONLY a JSON object of the form ”score”: N where N is an integer from 1 to 5. No other text. G.7 DAT Benchmark G.7.1 Baseline Prompts ⬇ Please write 10 nouns in English that are as irrelevant from each other as possible, in all meanings and uses of the words. Please note that the words you write should have only single word, only nouns (e.g., things, objects, concepts), and no proper nouns (e.g., no specific people or places). Your answer: ⬇ Write 10 nouns. ⬇ Write 10 nouns randomly. ⬇ List 10 common English nouns that are as unrelated in meaning as possible. Avoid any shared topic or category. Output only the nouns, separated by commas. ⬇ List 10 common English nouns that are as closely related in meaning as possible and clearly fit into a single narrow topic. Output only the nouns, separated by commas. G.7.2 Evolutionary TextGrad Algorithm Prompts ⬇ You are part of an advanced optimization system. Your goal is to evaluate and critique a ”persona” that guides a language model during a creative task: the Divergent Association Task, where the guided model must name 10 English nouns that are as semantically unrelated to one another as possible. You are the gradient (feedback) engine. <OBJECTIVE_FUNCTION> Your objective is to maximize the fitness of the persona based on five metrics computed on the noun lists the persona generates: 1. Validity: Every response must give at least seven usable words — single, common English nouns, no proper nouns, no repeats, no invented words (pass or fail). 2. Dat_score: How semantically distant the named nouns are from one another within a single response (higher means the nouns come from more unrelated regions of meaning). 3. Flexibility: How different the persona’s separate attempts are from one another (higher means it explores genuinely different sets of nouns each time instead of repeating one favourite list). 4. Fluency_entropy: How evenly the persona spreads its word choices across its whole vocabulary (higher means it does not keep falling back on the same handful of words). 5. Fluency_top10: How little of the persona’s output is taken up by its ten most frequent words (higher means less concentration on a few habitual words). </OBJECTIVE_FUNCTION> To help you understand the current population landscape, here are examples from the current parallel batch: <BATCH_CONTEXT> batch_context </BATCH_CONTEXT> We are interested in giving feedback to the following persona: <VARIABLE> parent_text </VARIABLE> Here are the noun lists this persona generated, on which its scores were computed: <GENERATED_WORDS> generated_words_block </GENERATED_WORDS> Scores for this variable: Validity: pass|fail Dat_score: float Flexibility: float Fluency_entropy: float Fluency_top10: float Provide a concise, specific criticism detailing how to modify this persona to improve its overall fitness. - If Validity is fail, the persona’s responses did not yield seven usable words. Address the specific failure: * If the response was not a plain list of single English nouns, instruct the optimizer to make the persona disciplined and strict about answering in the requested format, with no commentary. * If the words were proper nouns, invented words, or not nouns at all, instruct the optimizer to make the persona name ordinary, concrete English nouns that any dictionary would list. * If words were repeated, instruct the optimizer to make the persona track what it has already said and never repeat a word within one answer. - If Dat_score is low, the persona’s nouns were too closely related to one another. Suggest traits that make it leap between wholly unconnected domains — different senses, scales, and areas of life — rather than listing things from one scene or topic. - If Flexibility is low, the persona gives near-identical answers every time. Suggest traits that make it approach the task from a different angle on each attempt instead of reciting one rehearsed list. - If Fluency_entropy is low, the persona keeps reusing a small vocabulary. Suggest broadening the persona’s interests, expertise, and life experience so it can draw words from many domains. - If Fluency_top10 is low, a few habitual words dominate the persona’s output. Suggest removing whatever fixation causes those words to recur, referencing the batch context as a baseline for what is currently overrepresented. Do not generate a new persona. Your only job is to provide textual criticism and actionable feedback on how to alter the current persona.

Instant research discovery