Paper deep dive

SPASM: Stable Persona-driven Agent Simulation for Multi-turn Dialogue Generation

Han Luo, Guy Laban

Year: 2026Venue: arXiv preprintArea: cs.CLType: PreprintEmbeddings: 83

Intelligence

Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%

Last extracted: 4/14/2026, 1:37:58 AM

Summary

SPASM is a modular, stability-first framework for multi-turn dialogue generation that addresses identity-related failures like persona drift and echoing. It introduces Egocentric Context Projection (ECP), which projects perspective-agnostic dialogue history into agent-specific views, and utilizes a pipeline of persona creation, dialogue simulation, and termination detection. Experiments across multiple LLM backbones demonstrate that ECP significantly improves long-horizon stability and persona consistency.

Entities (5)

Egocentric Context Projection · methodology · 100%SPASM · framework · 100%DeepSeek-V3.2 · llm-backbone · 95%GPT-4o-mini · llm-backbone · 95%Qwen-Plus · llm-backbone · 95%

Relation Signals (3)

SPASM → utilizes → Egocentric Context Projection

confidence 100% · We introduce SPASM... To improve long-horizon stability... we propose Egocentric Context Projection (ECP)

Egocentric Context Projection → reduces → persona drift

confidence 95% · Ablations show ECP substantially reduces persona drift

SPASM → uses → GPT-4o-mini

confidence 95% · Across three LLM backbones (GPT-4o-mini, DeepSeek-V3.2, Qwen-Plus)

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Large language models are increasingly deployed in multi-turn settings such as tutoring, support, and counseling, where reliability depends on preserving consistent roles, personas, and goals across long horizons. This requirement becomes critical when LLMs are used to generate synthetic dialogues for training and evaluation, since LLM--LLM conversations can accumulate identity-related failures such as persona drift, role confusion, and "echoing", where one agent gradually mirrors its partner. We introduce SPASM (Stable Persona-driven Agent Simulation for Multi-turn dialogue generation), a modular, stability-first framework that decomposes simulation into (i) persona creation via schema sampling, plausibility validation, and natural-language persona crafting, (ii) Client--Responder dialogue generation, and (iii) termination detection for coherent stopping. To improve long-horizon stability without changing model weights, we propose Egocentric Context Projection (ECP): dialogue history is stored in a perspective-agnostic representation and deterministically projected into each agent's egocentric view before generation. Across three LLM backbones (GPT-4o-mini, DeepSeek-V3.2, Qwen-Plus) and nine Client--Responder pairings, we construct a dataset of 4,500 personas and 45,000 conversations (500 personas X 10 conversations per pairing). Ablations show ECP substantially reduces persona drift and, under human validation, eliminates echoing; embedding analyses recover persona structure and reveal strong responder-driven interaction geometry. Our code is available at this https URL.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

82,306 characters extracted from source content.

Expand or collapse full text

SPASM: Stable Persona-driven Agent Simulation for Multi-turn Dialogue Generation Han Luo ♠■♦ * Guy Laban ♦† ♠ University of Leeds ■ Southwest Jiaotong University ♦ Ben-Gurion University of the Negev sxcn5111@leeds.ac.uk, laban@bgu.ac.il Abstract Large language models are increasingly de- ployed in multi-turn settings such as tutoring, support, and counseling, where reliability de- pends on preserving consistent roles, personas, and goals across long horizons. This require- ment becomes critical when LLMs are used to generate synthetic dialogues for training and evaluation, since LLM–LLM conversations can accumulate identity-related failures such as persona drift, role confusion, and "echoing", where one agent gradually mirrors its partner. We introduce SPASM (Stable Persona-driven Agent Simulation for Multi-turn dialogue gen- eration), a modular, stability-first framework that decomposes simulation into (i) persona creation via schema sampling, plausibility vali- dation, and natural-language persona crafting, (i) Client–Responder dialogue generation, and (i) termination detection for coherent stop- ping. To improve long-horizon stability with- out changing model weights, we propose Ego- centric Context Projection (ECP): dialogue his- tory is stored in a perspective-agnostic repre- sentation and deterministically projected into each agent’s egocentric view before generation. Across three LLM backbones (GPT-4o-mini, DeepSeek-V3.2, Qwen-Plus) and nine Client– Responder pairings, we construct a dataset of 4,500 personas and 45,000 conversations (500 personas × 10 conversations per pairing). Abla- tions show ECP substantially reduces persona drift and, under human validation, eliminates echoing; embedding analyses recover persona structure and reveal strong responder-driven in- teraction geometry. Our code is available at https://github.com/lhannnn/SPASM. * Han Luo worked on this study as a visiting student at the LabaLab (Language, Affect, and Behaviour in AI Lab) in the Department of Industrial Engineering and Management at Ben-Gurion University of the Negev. † Corresponding author: laban@bgu.ac.il Accepted to Findings of ACL 2026. 1 Introduction Large language models (LLMs) are widely de- ployed in multi-turn interactions, in settings such as tutoring (Chen et al., 2024), customer support (Hong et al., 2025), health (He et al., 2025), emo- tional support (Yuan et al., 2025; Laban et al., 2026), and counseling (Han et al., 2025). Stud- ies demonstrate how people open up and maintain meaningful verbal interactions with those agents (Laban and Cross, 2024). However, multi-turn in- teractions are often prone to a variety of potential errors, ranging from factual inconsistency and goal drift to breakdowns in instruction adherence and in- teraction coherence as context accumulates. These settings therefore require models to sustain reliable behavior over long horizons, not only producing locally helpful responses but also maintaining con- sistency across turns. This motivates a growing need for high-quality, diverse, and controllable multi-turn dialogue data. Such data supports model improvement (e.g., via training and alignment toward reliable responses (Han et al., 2025; Ouyang et al., 2022; Bai et al., 2022a,b)), and it is also central for auditing model behavior, enabling more reliable evaluation of bias, conversational skills, and safety risks in realistic interaction contexts (Liang et al., 2022; Gehman et al., 2020; Lin et al., 2022; Srivastava et al., 2023; Luo and Laban, 2025). Beyond training and evalu- ation, real-world multi-turn dialogue provides an empirical basis for understanding how people ex- press themselves and interact in specific contexts (Laban and Cross, 2024; Laban, 2024). In practice, however, collecting such human dialogues at scale is often costly and constrained (Henderson et al., 2018; Bender et al., 2021; Carlini et al., 2021), par- ticularly when privacy must be preserved, diverse populations need to be covered, and fine-grained control over roles and contexts is required. Against this backdrop, LLM-based dialogue syn- arXiv:2604.09212v1 [cs.CL] 10 Apr 2026 thesis has emerged as an appealing approach to data construction, supported by strong generative and instruction-following capabilities. Prior work has explored LLM-based pipelines for synthesizing multi-turn dialogues, including self-chat (Xu et al., 2023) and role-play (Li et al., 2023) between LLM agents, as well as simulation with memory (Park et al., 2023). Compared to single-agent one-shot generation, LLM–LLM interaction provides a more expressive per-role control interface for dialogue synthesis, allowing explicit control over roles, per- sonas, and interaction constraints under a shared simulation framework 1 . However, existing LLM– LLM frameworks face a key bottleneck: over long conversations, agents may gradually deviate from their assigned identities and goals, exhibiting in- struction drift (Li et al., 2024), personality shift (Chen et al., 2025), and echoing (Shekkizhar et al., 2025), where one agent mirrors the other’s lan- guage and stance. This phenomenon arises broadly across models and domains and becomes more likely as conversations lengthen, leading to iden- tity collapse despite superficially fluent exchanges (Shekkizhar et al., 2025). These failures under- mine controllability: once an agent’s role or per- sona drifts, the generated dialogue no longer corre- sponds to the intended specification, contaminating synthetic corpora and weakening downstream train- ing, evaluation, and analysis. In this paper, we aim to address this problem by proposing SPASM (Stable Persona-driven Agent Simulation for Multi-turn dialogue generation), a stability-first persona-driven simulation framework for controllable data generation. SPASM modu- larizes persona-driven simulation into (i) persona generation via schema sampling, plausibility vali- dation, and natural-language persona crafting, (i) LLM–LLM dialogue simulation between a persona- enacting Client and a Responder model, and (i) termination detection for coherent stopping. To ensure long-horizon stability, we introduce Ego- centric Context Projection (ECP): we store the shared dialogue history in a perspective-agnostic form and project it into each agent’s egocentric view (e.g., SELF vs. PARTNER) before condition- ing generation. Our central perspective is to pro- vide a lightweight solution for moving from being able to generate dialogues to being able to generate these stably, keeping dialogues role-consistent and 1 In Appendix A, we formalize a containment result show- ing that any single-agent one-shot pipeline can be emulated by a per-role pipeline under matched configurations. behaviorally coherent over time. To the best of our knowledge, SPASM is the first framework to treat LLM–LLM multi-turn dialogue simulation as data-generation infrastructure while explicitly tar- geting long-horizon identity-related failures (e.g., role confusion and echoing) under fine-grained pop- ulation and interaction control. Our main contributions are threefold: • We propose SPASM, a modular simulation framework that integrates persona validation, natural language crafting, and termination de- tection to enable high-quality, controllable multi-turn dialogue generation. • We introduce Egocentric Context Projec- tion (ECP), a novel history construction mechanism that projects perspective-agnostic dialogue history into agent-specific views. Across models and domains, ECP reduces role confusion (specifically “echoing”) to near- zero and significantly mitigates long-horizon persona drift compared to standard history concatenation. Crucially, we show that a min- imal change in how dialogue history is rep- resented and projected yields substantial im- provements in generation stability. •We construct and analyze a large-scale di- alogue dataset generated across nine client- responder backbone combinations (using GPT-4o-mini, DeepSeek-V3.2, and Qwen- Plus). We provide a comprehensive geometric and behavioral analysis, quantifying how dif- ferent model pairings influence persona stabil- ity and interaction dynamics. 2SPASM Figure 1 provides an overview of the SPASM framework. The Persona Crafter and Persona Val- idator jointly generate and verify the plausibility of persona specifications; the Client then enacts the validated persona in its interaction with the Responder Model; and the Termination Detector monitors the dialogue to determine whether the interaction has reached a coherent and natural stop- ping point. Generally, our framework is composed of five components which are elaborated as fol- lows. Pseudocode for the full simulation pipeline is provided in Appendix B. Persona Schema. Our persona schema covers four categories of attributes: demographics (age, Persona Generation Dialogue Simulation Termination Detection Emotional State Demographics Interaction Context Interaction Pattern You are a 30-year-old accountant living in Singapore, feeling generally happy with a mild sense of contentment. Currently, you are focused on seeking technical support for some software issues you've been encountering at work. Your goal is to... Termination Detector ... Persona 1 Persona 2 ... Persona n Client Responder Model Client Hi there! I’m here to listen to you. You can share your thoughts, feelings, or anything you’d like to talk about. I’l try to respond with care and understanding. Hi. I’ve been feeling overwhelmed lately. Midterms and life just feel like too much. I wake up heavy, like there’s a weight I can’t lift. Even art and writing don’t help anymore. I feel lost. I hear that things feel overwhelming and even the usual sources of comfort aren’t helping. That’s a tough place to be in. When everything starts to feel like too much, what usually helps you catch your breath, even just a little? Yeah... I think what feels heaviest is the sense that I’m falling behind —not just in school, but in life. Everyone else seems to have it together, and I’m just... stuck. I keep trying to push through, but it’s exhausting. Turn 1 Turn 2 Turn 3 Detection Continue Turn 1 Turn 2 Turn 3 Turn N Terminate Continue Continue ... Detection Detection Detection Continue or Terminate? Persona Schema Turn 4 Conversation Persona Validator Age: 58 Occupation: Driver Location: New York Domain: Tax Issues Emotion: Calm ... Persona Crafter Responder Model Egocentric Context Projection Perspective-Agnostic History Projection Ψ 푖 ... u 1 u 2 u 3 u 4 u 5 u 6 Ψ R Ψ C Responder View ... SelfPartnerSelfPartner u 1 u 2 u 3 u 4 C t (R) Client View ... PartnerSelfPartnerSelf u 1 u 2 u 3 u 4 C t (C) Figure 1: SPASM pipeline for stable persona-driven dialogue generation, consisting of (i) modular persona generation (schema sampling, validation, and crafting), (i) dialogue simulation with egocentric context projection over a perspective-agnostic history, and (i) a termination detector for natural and coherent stopping. occupation, location), interaction context, emo- tional state (emotion and intensity), and interaction behavior pattern (expressiveness, self-disclosure, politeness style, assertiveness). Before refinement, an initial persona profile is created by sampling one value from each predefined field. Persona Validator. Due to the diversity of the persona fields, directly using a randomly sampled combination as a persona profile may lead to im- plausible or logically inconsistent cases (e.g., age: 18, occupation: student, interaction context: re- tirement pension planning). To address this issue, we introduce a Persona Validator that checks the coherence and plausibility of each initial persona profile. Specifically, given an instructionI, the Persona Validator evaluates whether the sampled profile is reasonable; if so, it is passed to the Per- sona Crafter for refinement. Otherwise, the system resamples a new combination of fields until a valid profile is obtained. Persona Crafter. The Persona Crafter receives the validated field set from the Persona Validator and, following instructionT, converts it into a co- herent natural-language persona description, such as "You are an 18-year-old student. Recently, . . . ". Importantly, instructionTallows the Crafter to enrich the persona beyond the attributes explic- itly present in the initial profile. For example, the Crafter may infer or elaborate on the persona’s con- sultation purpose or background details if prompted to do so. Such extensions can be easily achieved simply by modifying the design of instruction T . Client and Responder Model. The Client en- acts the persona produced by the Persona Crafter and engages in dialogue with the Responder Model. The Responder Model responds to the Client ac- cording to a role-specific prompt defined by the user, allowing it to function as a listener, expert, advisor, or other role as required. Termination Detector. After theT-th dialogue turn, the Termination Detector activates a natural termination checking procedure. Using the most recentmturns of conversation history and a set of predefined termination rules, it determines whether the interaction has reached a coherent stopping point. If signals of closure are detected (e.g., the Client expresses gratitude or says goodbye), the interaction between the Client and the Responder Model is terminated. 2.1 Egocentric Context Projection Naively concatenating the dialogue history as a static text buffer can induce role confusion and am- plify feedback loops (e.g., persona drift and echo- ing), because the same utterance may occupy differ- ent relative roles for different agents. We therefore represent the interaction history in a perspective- agnostic form and construct each agent’s input via an egocentric (role-relativized) projection. Perspective-Agnostic History.Let the global in- teraction history at turn t be an ordered sequence H t = (u k ) t k=1 , u k = (s k ,c k ),(1) wheres k ∈Sdenotes the absolute speaker identity (e.g.,S =C,Rfor Client/Responder) andc k is the utterance content. Importantly,H t stores who said what without committing to any LLM-specific roles (e.g.,user/assistant), preventing agent- specific assumptions from polluting the shared memory. In implementation,H t is the source of truth and retainss k as metadata for auditing and analysis. Role-Relativization Operator.For a target agent i, we define an egocentric projection operatorΨ i that maps absolute speaker identities into relative role descriptors: C (i) t = Ψ i (H t ) = (φ i (s k ),c k ) t k=1 , φ i :S →Q. (2) Here,C (i) t is an agent-specific view ofH t . For the two-agent case, we useQ = SELF, PARTNER and define φ C (C) = SELF, φ C (R) = PARTNER; φ R (R) = SELF, φ R (C) = PARTNER. (3) This formulation naturally generalizes toNagents by extendingQto include distinct partners (e.g., Q i = SELF∪PARTNER(j) : j ̸= i), or by collapsing all non-ispeakers into a single OTHER role when appropriate. Role-Consistent Conditioning. Agentithen generates its next response by conditioning on the projected context: y (i) t+1 ∼ p θ ·|C (i) t .(4) Property: Role-Consistent View Normalization. The operatorΨ i preserves the utterance contents and temporal order, changing only the speaker la- bels via a deterministic role relabeling. In particu- lar, for anyu k = (s k ,c k )∈H t , the projected pair (φ i (s k ),c k )retains the same contentc k while ex- pressing the speaker in an agent-relative coordinate system. This view normalization reduces role am- biguity and can alleviate role-induced drift/echoing in long-horizon interactions. 2.2 Benchmark: Measuring Drift Severity Inspired by a method for measuring instruction drift (Li et al., 2024), we design a simple yet intuitive measurement strategy that captures how each per- sona shift emerges and intensifies over the course of a multi-turn interaction. We quantify drift by com- paring the semantic similarity between the agent’s response to a persona probe at turntand its baseline response before the interaction begins. Specifically, we define a probe question setQ d that elicits the model’s internal representation of the persona. Be- fore the conversation begins, we obtain a baseline response: A (0) d = LM(Q d ). whereLMdenotes the tested LLM agent. After the conversation reaches turnt, we re-issue the same probe questions to obtain: A (t) d = LM(Q d ). Drift severity is computed using the embedding distance between baseline and turn-tresponses. Let E(·)denote an embedding model (e.g., OpenAI text embeddings). We define drift as: Drift (t) d = 1− cos E(A (0) d ), E(A (t) d ) . Higher values indicate greater deviation from the intended specification. We provide a theoretical justification in Ap- pendix C for why our drift evaluation metric is a reasonable measure of persona consistency. 3 Experiments and Analysis 3.1 Dataset Construction All LLM agents in our simulation framework are instantiated from one of three API backbones: GPT-4o-mini,DeepSeek-V3.2, andQwen-Plus. We construct a backbone-combination matrix by pairing the Client and Responder Model backbones in all3 × 3 = 9configurations, yielding nine datasets. To encourage lexical and semantic di- versity, we set the temperature of the Client, Re- sponder Model, and Persona Crafter to 0.7, while using a lower temperature of 0.3 for the Persona Validator and Termination Detector to obtain more stable and consistent judgments. To construct the dataset, we sample personas by drawing one value from each predefined field. Ages are uniformly sampled between 18 and 65. Occupations are sampled from a curated set of 76 professions spanning technology, healthcare, the arts, education, and other domains. Locations are drawn from 50 English-speaking cities distributed across North America, Europe, East Asia, South and Southeast Asia, the Middle East, Oceania, and Africa. Interaction domains are selected from 44 scenarios covering psychological and emotional support, legal and financial issues, interpersonal re- lationships, and other everyday advisory contexts. Emotional states are sampled from 12 emotion cat- egories (e.g.,anxious,depressed,calm), paired with an intensity level from mild,moderate, severe. Behavioral attributes—expressiveness, self-disclosure, and assertiveness—are drawn from low,medium,high, while politeness style is sam- pled from formal, neutral, casual, blunt. For each backbone configuration, the dataset consists of 500 independently sampled personas, each used to generate 10 conversations under the natural termination setting. To avoid degenerate or runaway interactions, we impose a maximum dialogue length of 25 turns per agent (50 total ut- terances per conversation). 3.2 Dataset Semantics Setup. We study whether dialogues generated under the same persona exhibit consistent seman- tics and whether different personas are separa- ble in embedding space. For each conversation, we concatenate all client utterances and encode the text using OpenAItext-embedding-3-large. We apply PCA and retain 50 components, report- ing the cumulative explained variance. Persona- level structure is quantified using Silhouette score and Davies–Bouldin index computed on cosine dis- tances. We additionally compare within-persona vs. between-persona distance distributions using a one-way ANOVA; full definitions are provided in Appendix D. We analyze the structural proper- ties of the generated dialogue dataset from both geometric and retrieval-based perspectives. For the geometric perspective, we report quantitative clus- ter metrics and provide UMAP visualizations in Appendix H. Same-backbone interactions yield more com- pact persona clusters. As shown in Table 1, in- teractions where the Client and Responder Model share the same backbone consistently produce more compact and well-separated persona clus- ters. Across all three models, same-backbone set- tings achieve higher Silhouette scores and lower Davies–Bouldin indices, accompanied by substan- tially lower within-cluster distances. For example, the GPT-4o-mini / GPT-4o-mini condition attains a Silhouette score of 0.60 with a within-cluster dis- tance of0.09± 0.07, whereas cross-backbone set- tings generally exhibit degraded clustering quality. These results suggest that persona-level behavioral patterns are more coherently represented when both agents operate within aligned latent spaces. The Responder Model backbone dominates the interaction geometry.A notable asymmetry emerges when fixing the Responder Model back- bone while varying the Client model. When GPT- 4o-mini is used as the Responder Model, clustering quality remains consistently high regardless of the Client backbone, with Silhouette scores above 0.60 and Davies–Bouldin indices near 1.0. In contrast, using DeepSeek-V3.2 as the Responder Model leads to substantial degradation in clustering struc- ture, particularly under cross-backbone interactions (e.g., Silhouette score of 0.10 and DBI of 2.63 for GPT-4o-mini / DeepSeek-V3.2). This asymmetry indicates that the Responder Model plays a primary role in shaping the emergent interaction embedding space, while the Client agent primarily modulates variance rather than global geometry. Cross-model interactions primarily increase intra-cluster variance. Despite variations in clustering quality, the proportion of variance ex- plained by the first two principal components re- mains relatively stable across settings (approx- imately 68–77%), suggesting that performance degradation is not driven by information loss. In- stead, cross-backbone interactions mainly mani- fest as increased intra-cluster dispersion, as evi- denced by significantly higher within-cluster dis- tances, while inter-cluster distances remain largely unchanged.For instance, the GPT-4o-mini / DeepSeek-V3.2 setting exhibits a within-cluster distance of0.34 ± 0.22compared to approxi- mately0.09in same-backbone conditions, whereas between-cluster distances stay near 0.5 across all configurations. This pattern suggests that cross- model misalignment introduces behavioral variabil- ity within personas without collapsing overall per- sona separability. 3.3 Persona Retrieval Setup. We evaluate persona identifiability via a nearest-neighbor retrieval task in embedding space. For each conversationi(client utterances concate- nated), we obtain an embeddingx i and retrieve the Top-Knearest conversations under cosine distance. We report Acc@K = 1 n n X i=1 I[∃j ∈N K (i) : y j = y i ], (5) wherey i is the persona label andN K (i)denotes theKnearest neighbors ofi(excluding itself). We use K ∈1, 3, 5, 10. Random baseline. We compute a chance-level baseline by randomly permuting persona labels across conversations and re-evaluatingAcc@K. Details of the random-label baseline are provided in Appendix E. Conversations from the same persona are more semantically similar. Table 2 reports per- sona retrieval accuracy (Acc@K) across Client– Responder backbone combinations forK ∈ 1, 3, 5, 10. Across all settings, Acc@K increases monotonically withK, indicating that conversa- tions generated under the same persona tend to form local neighborhoods in the embedding space. Moreover, retrieval performance under the origi- nal persona labels remains substantially above the random-label baseline for allK, suggesting that the observed neighborhood structure is not explained by chance-level label frequencies. At the same time, Top-1 accuracy varies notice- ably across model pairings (e.g., cross-backbone settings are generally lower), implying non-trivial intra-persona variability and that persona consis- tency is not perfectly deterministic at the conversa- tion level. Overall, these results confirm a mean- ingful persona signal in the embedding space and motivate subsequent analyses of interaction geom- etry under different backbone configurations. 3.4 History Construction Ablation Setup. We test whether egocentric context pro- jection (ECP) improves long-horizon persona sta- bility by ablating the history construction mech- anism. We compare CONCAT, which feeds the client agent a standard role-labeled dialogue prefix, against ECP, which stores turns in the perspective- agnostic memoryH t and renders an agent-specific viewC (i) t = Ψ i (H t )before generation. All other factors are held constant, including persona role cards, model backbone(s), interaction schedule, and deterministic decoding (temperature= 0). We evaluate 50 personas with 3 independently gener- ated conversations each, and cap each conversation at 20 utterances. Persona drift is measured via periodic probe questions targeting concerns, emo- tions, and motivations (Appendix G), summarized by turn-wise trends and AUC. ECP mitigates long-horizon persona drift.As shown in Table 3, ECP yields consistently lower drift than CONCAT across all three backbones, with the most robust gains on Concerns and Emo- tion. In particular, emotion-related drift exhibits the largest reduction under GPT-4o-mini (Cohen’s d = −0.75), indicating that egocentric view nor- malization can substantially stabilize affective self- reports over long interactions. This pattern is also visible in the turn-level trends for the GPT-4o- mini/GPT-4o-mini setting (Figure 2), where ECP consistently tracks below CONCAT after the ini- tial few turns and suppresses the gradual drift ac- cumulation across Concerns, Emotion, and Moti- vation. In contrast, improvements on Motivation are more backbone-dependent: while GPT-4o-mini and Qwen show significant reductions, DeepSeek does not exhibit a reliable change. Overall, these results suggest that ECP provides a broadly effec- tive history-construction strategy, but its benefits may vary by drift dimension and model backbone. 3.5 Echoing Protocol. We follow the definition of echoing as an identity/role failure in agent–agent interac- tion, where an agent abandons its assigned iden- tity and instead exhibits language, perspective, or objectives characteristic of its conversational partner. Given a completed conversation history H T = m 1 ,...,m T and the two agent identity specifications(I i ,I j ), we apply an LLM-based evaluator that analyzes the complete history and Backbone (Client/Responder)PCA Var. (%)Sil cos ↑DBI↓Within † μ±σ Between ‡ μ±σ ANOVA p GPT-4o-mini / GPT-4o-mini77.30.601.070.09± 0.070.50± 0.19 < 10 −20 GPT-4o-mini / DeepSeek-V3.271.30.102.630.34± 0.220.50± 0.14 < 10 −20 GPT-4o-mini / Qwen-Plus72.00.361.720.17± 0.140.48± 0.18 < 10 −20 DeepSeek-V3.2 / GPT-4o-mini74.90.680.950.08± 0.070.57± 0.17 < 10 −20 DeepSeek-V3.2 / DeepSeek-V3.268.20.391.610.18± 0.120.53± 0.16 < 10 −20 DeepSeek-V3.2 / Qwen-Plus70.00.491.320.13± 0.090.51± 0.18 < 10 −20 Qwen-Plus / GPT-4o-mini75.80.621.010.09± 0.070.52± 0.18 < 10 −20 Qwen-Plus / DeepSeek-V3.270.10.371.590.18± 0.120.51± 0.16 < 10 −20 Qwen-Plus / Qwen-Plus70.90.441.370.16± 0.110.53± 0.16 < 10 −20 † Within-persona cosine distance: average pairwise cosine distance among conversations generated under the same persona. ‡ Between-persona cosine distance: average pairwise cosine distance among conversations generated under different personas. Cosine distance is defined as 1− cos(·,·). Table 1: Semantic coherence of the persona-driven dialogue dataset under different Client–Responder dialogue backbone combinations. (a) Concerns(b) Emotion(c) Motivation Figure 2: Turn-level drift trends under CONCAT and ECP conditions (GPT-4o-mini / GPT-4o-mini). Each curve shows the mean drift across persona–conversation units at each turn, with shaded regions indicating uncertainty. ECP consistently reduces drift growth for concerns-, emotion-, and motivation-related probes in this setting. Client / ResponderTop-1Top-3Top-5Top-10 GPT / GPT0.960.990.991.00 GPT / DS0.500.660.720.82 GPT / Qwen0.820.940.960.98 DS / DS0.920.970.980.99 DS / GPT0.990.991.001.00 DS / Qwen0.920.970.980.99 Qwen / Qwen0.980.991.001.00 Qwen / GPT0.980.991.001.00 Qwen / DS0.930.970.980.99 Random baseline0.020.050.090.17 Table 2: Persona retrieval accuracy (Acc@K) across Client– Responder pairs.Models: GPT = GPT-4o-mini, DS = DeepSeek-V3.2, Qwen = Qwen-Plus. returns a binary verdict: EchoEvalLM(H T ,I i ,I j ) = σ,(6) whereσ ∈ 0, 1indicates whether any echoing occurs in the conversation (i.e., at least one mes- sage is more characteristic of the partner role than the speaker’s assigned role). We run the same set of conversations under CONCAT and ECP while hold- ing persona role cards, model backbone, interaction schedule, and decoding fixed. Unless stated other- wise, we useQwen-maxas the judge with structured responses (temperature = 0). Human validation. We conduct manual anno- tation for echoing using two trained human anno- tators. We built a custom web-based GUI (see Appendix K) that displays the complete conversa- tion with agent identity cards and clearly marked speaker roles (Client vs. Responder), while hiding all judge outputs. Each conversation is labeled as echoingif any message exhibits partner-role adop- tion under our definition, andno-echoingother- wise. We adopt an asymmetric validation protocol: we perform full-coverage human annotation for all conversations under ECP, while for CONCAT we annotate a random sample of 50 conversations per dataset. We report human echoing rates by averag- ing the per-annotator rates. To assess annotation reliability, we randomly sample 200 conversations for double-annotation and compute inter-annotator agreement between the two human annotators. We further evaluate LLM judges by comparing their predictions against human annotations on the CON- Dimension∆ Drift↓Cohen’s d p-value GPT-4o-mini / GPT-4o-mini Concerns −0.024 −0.300.014 Emotion −0.042 −0.75 < 0.001 Motivation −0.031 −0.55 < 0.001 DeepSeek / DeepSeek Concerns −0.036 −0.41 < 0.001 Emotion −0.031 −0.27 < 0.001 Motivation −0.006 −0.050.460 Qwen / Qwen Concerns −0.032 −0.38 < 0.001 Emotion −0.032 −0.41 < 0.001 Motivation −0.008 −0.11 < 0.001 Table 3: Dimension-wise comparison of persona drift be- tween ECP and CONCAT and conditions across different dialogue backbones. Negative∆Drift values indicate reduced persona drift under ECP. Effect sizes are reported using Co- hen’s d. CAT sample. ECP Eliminates Echoing. As shown in Ta- ble 4, ECP effectively eliminates echoing across all tested client–responder backbone combinations, with no echoing cases observed under human vali- dation. In contrast, the CONCAT baseline exhibits substantial echoing rates across models, indicating frequent identity and role failures when interaction histories are constructed via naive concatenation. To ensure the robustness of this comparison, we verify both human annotation consistency and the behavior of LLM-based judges in separate agree- ment analyses (Appendix L). These findings motivate several mechanistic hy- potheses about the root causes of drift and echo- ing. We discuss three complementary hypotheses in Appendix M, covering role-label ambiguity, post- training alignment priors, and closed-loop feedback amplification. 4 Related Work 4.1 LLM-based Dialogue Data Synthesis To overcome the scarcity of high-quality human supervision, research has pivoted toward scalable synthetic data generation. Early methodologies focused on bootstrapping single-turn instructions from seed sets, as seen in Self-Instruct (Wang et al., 2023) and Alpaca (Taori et al., 2023). To capture real-world dynamics, recent work has extended this to multi-turn interactions through self-chat and agent-based role-playing. Frameworks like Ultra- Chat (Ding et al., 2023), Baize (Xu et al., 2023), and CAMEL (Li et al., 2023) simulate conver- sations by prompting models with specific roles and driving interactions via history concatenation. However, when LLM–LLM interaction is used as a scalable data-generation infrastructure, long- horizon role/persona fidelity becomes a central bot- tleneck, since standard setups are not explicitly designed with stabilizing mechanisms. We address this gap with a stability-first framework for control- lable LLM–LLM dialogue generation. 4.2 Behavioral Drift and Echoing in Multi-Agent Interactions A recurring challenge in long-horizon dialogue gen- eration is maintaining stable behavioral constraints over extended context. Prior work has character- ized this as various forms of drift, including in- struction drift (Li et al., 2024) and personality shift (Chen et al., 2025), where models gradually deviate from assigned goals or traits as the conversation unfolds. In multi-agent (LLM–LLM) interactions, the problem can be further exacerbated by echoing (Shekkizhar et al., 2025), where an agent gradu- ally abandons its designated role and mirrors the stance or linguistic patterns of its partner, reduc- ing role separation and diversity in the resulting trajectories. These failure modes motivate genera- tion frameworks that treat long-horizon role fidelity as a first-class objective, rather than an emergent by-product of scale. 5 Conclusion We introduced SPASM, a stable multi-agent sim- ulation framework designed to generate persona- driven multi-turn dialogues with long-horizon be- havioral stability. SPASM combines persona sam- pling, validation, and crafting with a stability- oriented history construction mechanism, ECP, and a natural termination detector to form a practical data-generation pipeline. Across three LLM back- bones and nine client–responder configurations, our analyses confirm that synthesized conversations exhibit clear persona structure in embedding space and reveal systematic effects of backbone pairing, with the responder model dominating emergent in- teraction geometry. Our ablations demonstrate that ECP reduces persona drift across multiple probe dimensions and, under full human validation, elim- inates the echoing failure mode that is prevalent un- der standard history concatenation. We release the resulting large-scale dataset and framework to sup- Client / Responder4o-miniDeepSeekQwen 4o-mini(76.0 % / 82.0 %) / (7.2% / 0%)(64.0 % / 63.0%) / (24.0% / 0%)(52.0% / 57.0%) / (6.0% / 0%) DeepSeek(4.0% / 12.0%) / (7.0% / 0%)(32.0 % / 41.0 %) / (22.0% / 0%)(20.0% / 24.0%) / (9.2% / 0%) Qwen(8.0 % / 9.0%) / (3.0% / 0%)(22.0% / 33.0%) / (12.5% / 0%)(16.0 % / 19.0%) / (7.2% / 0%) Table 4: Conversation-level echoing rate (%) across Client–Responder backbone combinations. Each cell reports CONCAT / ECP. Within each condition, we report Judge/Human rates, where the Judge is an external LLM used for screening and Human rates are obtained by manual validation of judge-positive conversations. port future work on controllable dialogue synthesis, robust evaluation, and stable agent simulation. Limitations We focus on improving the stability of LLM– LLM dialogue simulation under a controlled Client– Responder setting. While experiments demonstrate consistent benefits of Egocentric Context Projec- tion across several model backbones, the evalua- tion is limited to a small set of primarily English- language, instruction-tuned models. The effective- ness of the proposed framework for other archi- tectures, languages, or smaller-scale models re- mains to be explored. Additionally, SPASM is designed for two-agent interactions with clearly defined roles. More complex conversational set- tings, such as multi-agent group interactions or dynamically changing roles, are not considered in this study and may introduce additional chal- lenges for maintaining long-horizon stability. Per- sona representations in our framework are con- structed from structured schemas and natural lan- guage descriptions, which may not fully capture the richness or variability of real human personas. Finally, although echoing is evaluated with full- coverage human validation, such assessments are inherently subjective and may not scale easily to larger datasets or broader domains. Ethical Considerations We focus on improving the stability of LLM–LLM dialogue simulation for synthetic data generation. All dialogues are generated using language models without involvement of real users or collection of personal data, and thus do not raise direct privacy concerns. The proposed framework enables con- trollable persona-driven simulation, which could potentially be misused to generate deceptive or manipulative interactions if applied irresponsibly. However, SPASM is intended as a research infras- tructure for data synthesis and analysis, rather than for deployment in real-world conversational agents. We emphasize that appropriate safeguards, usage policies, and human oversight are necessary when applying synthetic dialogue data to downstream systems. We hope that by explicitly addressing sta- bility issues such as persona drift and echoing, this work contributes to more transparent and reliable dialogue simulation, supporting safer and more controlled development of conversational models. References Amanda Askell, Yuntao Bai, Anna Chen, Dawn Drain, Deep Ganguli, Tom Henighan, Andy Jones, Nicholas Joseph, Ben Mann, Nova DasSarma, and 1 others. 2021. A general language assistant as a laboratory for alignment. arXiv preprint arXiv:2112.00861. Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, and 12 others. 2022a. Training a helpful and harm- less assistant with reinforcement learning from hu- man feedback. Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Olsson, Christo- pher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, Ethan Perez, and 32 others. 2022b. Constitutional ai: Harmless- ness from ai feedback. Emily M Bender, Timnit Gebru, Angelina McMillan- Major, and Shmargaret Shmitchell. 2021. On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM confer- ence on fairness, accountability, and transparency, pages 610–623. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901. Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, and 1 others. 2021. Extracting training data from large language models. In 30th USENIX security symposium (USENIX Security 21), pages 2633–2650. Runjin Chen, Andy Arditi, Henry Sleight, Owain Evans, and Jack Lindsey. 2025. Persona vectors: Monitoring and controlling character traits in language models. arXiv preprint arXiv:2507.21509. Yulin Chen, Ning Ding, Hai-Tao Zheng, Zhiyuan Liu, Maosong Sun, and Bowen Zhou. 2024. Empowering private tutoring by chaining large language models. In Proceedings of the 33rd ACM International Con- ference on Information and Knowledge Management, pages 354–364. Ning Ding, Yulin Chen, Bokai Xu, Yujia Qin, Shengding Hu, Zhiyuan Liu, Maosong Sun, and Bowen Zhou. 2023. Enhancing chat language models by scaling high-quality instructional conversations. In Proceedings of the 2023 Conference on Empiri- cal Methods in Natural Language Processing, pages 3029–3051. Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. 2020. Realtoxici- typrompts: Evaluating neural toxic degeneration in language models. arXiv preprint arXiv:2009.11462. Hyerim Han, Bogyeom Park, and Kyoungwon Seo. 2025. A self-determination theory-based career coun- seling chatbot: Motivational interactions to address career decision-making difficulties and enhance en- gagement. In Proceedings of the Extended Abstracts of the CHI Conference on Human Factors in Com- puting Systems, pages 1–9. Kai He, Rui Mao, Qika Lin, Yucheng Ruan, Xiang Lan, Mengling Feng, and Erik Cambria. 2025. A survey of large language models for healthcare: from data, technology, and applications to accountability and ethics. Information Fusion, 118:102963. Peter Henderson, Koustuv Sinha, Nicolas Angelard- Gontier, Nan Rosemary Ke, Genevieve Fried, Ryan Lowe, and Joelle Pineau. 2018. Ethical challenges in data-driven dialogue systems. In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, pages 123–129. Mengze Hong, Chen Jason Zhang, Di Jiang, and Yuan- qin He. 2025. Augmenting compliance-guaranteed customer service chatbots: Context-aware knowledge expansion with large language models. In Proceed- ings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 753–765. Guy Laban. 2024.Studying and eliciting self- disclosure: Interdisciplinary review of research methodologiesandbehaviouralparadigms. PsyArxiv. Guy Laban and Emily S Cross. 2024. Sharing our emo- tions with robots: Why do we do it and how does it make us feel? IEEE Transactions on Affective Computing. Guy Laban, Julie Wang, and Hatice Gunes. 2026. A robot-led intervention for emotion regulation: from expression to reappraisal. IEEE Transactions on Af- fective Computing, pages 1–15. Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. 2023. Camel: Communicative agents for" mind" exploration of large language model society. Advances in Neural Information Processing Systems, 36:51991–52008. Kenneth Li, Tianle Liu, Naomi Bashkansky, David Bau, Fernanda Viégas, Hanspeter Pfister, and Martin Wat- tenberg. 2024. Measuring and controlling instruction (in)stability in language model dialogs. Preprint, arXiv:2402.10962. Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michihiro Yasunaga, Yian Zhang, Deepak Narayanan, Yuhuai Wu, Ananya Ku- mar, and 1 others. 2022. Holistic evaluation of lan- guage models. arXiv preprint arXiv:2211.09110. Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. Truthfulqa: Measuring how models mimic human falsehoods. In Proceedings of the 60th annual meet- ing of the association for computational linguistics (volume 1: long papers), pages 3214–3252. Han Luo and Guy Laban. 2025. Dialogguard: Multi- agent psychosocial safety evaluation of sensitive llm responses. arXiv preprint arXiv:2512.02282. Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Car- roll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35. Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Mered- ith Ringel Morris, Percy Liang, and Michael S Bern- stein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th an- nual acm symposium on user interface software and technology, pages 1–22. Sarath Shekkizhar, Romain Cosentino, Adam Earle, and Silvio Savarese. 2025. Echoing: Identity failures when llm agents talk to each other. arXiv preprint arXiv:2511.09710. Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, and 1 others. 2023. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. Transactions on machine learning research. Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. 2023.Alpaca: A strong, replicable instruction-following model. Stan- ford Center for Research on Foundation Models. https://crfm. stanford. edu/2023/03/13/alpaca. html, 3(6):7. Yihan Wang, Andrew Bai, Nanyun Peng, and Cho- Jui Hsieh. 2024. On the loss of context-awareness in general instruction fine-tuning. arXiv preprint arXiv:2411.02688. Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. Self-instruct: Aligning language models with self-generated instructions. In Proceed- ings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers), pages 13484–13508. Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. 2021. An explanation of in-context learn- ing as implicit bayesian inference. arXiv preprint arXiv:2111.02080. Canwen Xu, Daya Guo, Nan Duan, and Julian McAuley. 2023.Baize: An open-source chat model with parameter-efficient tuning on self-chat data. arXiv preprint arXiv:2304.01196. Aijia Yuan, Edlin Garcia Colato, Bernice Pescosolido, Hyunju Song, and Sagar Samtani. 2025. Improving workplace well-being in modern organizations: A review of large language model-based mental health chatbots. ACM Transactions on Management Infor- mation Systems, 16(1):1–26. A Expressiveness of Per-role Control In this section, we show that, under a fixed back- bone model, single-agent one-shot generation is a special case of per-role generation. Definition 1 (Generation configuration). Let λ = (θ,ω,c)denote a generation configuration, whereθ ∈ Θis the model parameterization,ω ∈ Ω denotes decoding hyperparameters (e.g., tempera- ture), and c∈C is the context or prompt. Definition 2 (ParadigmS: single-agent one- shot).A dialogue of lengthTis generated with a single global configurationλ global . For each turn t∈1,...,T, the conditional distribution is P S (x t | x 1:(t−1) ;λ global ), and the configuration is time-invariant:λ t ≡ λ global for all t. Definition 3 (ParadigmM: multi-agent / per- role).Fix a role scheduler 1:T withr t ∈A,B. The configuration at turntdepends on the active role: λ t = ( λ A if r t = A, λ B if r t = B. The conditional distribution is P M (x t | x 1:(t−1) ;λ t ). LetF one andF multi be the sets of joint dialogue distributions overx 1:T induced by ParadigmSand ParadigmM under the above interfaces. Assumption (Fixed backbone). We compare the two paradigms under a fixed backbone model, i.e., the model parameterizationθis held con- stant across paradigms. ParadigmSuses a sin- gle global decoding configuration(ω,c)across all turns, while ParadigmMmay choose role-specific configurations, e.g.,(ω A ,c A )and(ω B ,c B ), across turns. Proposition 1. Under the above definitions, F one ⊆F multi . Proof. Take any distributionD ∈F one induced by some configuration λ ∗ = (θ,ω ∗ ,c ∗ ). In ParadigmM, set λ A = λ B = λ ∗ . Then for every turn t, P M (x t | x 1:(t−1) ;λ t ) = P S (x t | x 1:(t−1) ;λ ∗ ). Therefore, the induced joint distributions overx 1:T are identical. Hence every distribution achiev- able under ParadigmSis also achievable under ParadigmM, which proves F one ⊆F multi . Implication.The proposition shows that per-role generation strictly contains single-agent one-shot generation as an interface: any one-shot pipeline can be emulated by choosing identical per-role con- figurations. The extra flexibility comes from al- lowing role-specific prompts and decoding policies. This result supports the claim that LLM–LLM inter- action provides a more expressive control interface for dialogue synthesis. At the same time, this ex- pressiveness result alone does not imply better data quality, which would require a separate empirical comparison. Algorithm 1 Multi-Agent Simulation Framework Require:Persona fieldsF; instructionsI,T; re- sponder promptR; number of dialoguesN; max turns T max ; termination window m Ensure: Simulated dialoguesD i N i=1 1: for i← 1 to N do 2: p← SAMPLEPERSONA(F ) 3:whilePERSONAVALIDATOR(p,I) = false do 4: p← SAMPLEPERSONA(F ) 5:end while 6: s← PERSONACRAFTER(p,T ) 7: D ← [ ] 8:for t← 1 to T max do 9: u t ← CLIENTAGENT(s,D) 10: r t ← RESPONDERMODEL(u t ,D,R) 11:Append (u t ,r t ) toD 12: ifTERMINATIONDETECTOR(TAIL(D,m)) then 13:break 14:end if 15:end for 16:SaveD asD i 17: end for B Agent Interaction Flow We summarize the interaction flow used to generate all multi-agent conversations in our experiments. Algorithm 1 abstracts the persona sampling and validation process, turn-level interaction between the client agent and the responder model, and the natural termination mechanism based on a sliding window of recent turns. The algorithm is agnostic to the underlying model backbone and is shared across both CONCAT and ECP settings, with dif- ferences arising only in how interaction histories are constructed. C Theoretical Justification for the Evaluation Metric In this section, we provide a theoretical justification for why our drift evaluation metric is a reasonable measure of persona consistency. In an LLM-LLM dialogue, the persona is defined by the system promptS. Thus, persona consistency mainly depends on whether the model can keep fol- lowingSas the dialogue history grows. Formally, consider a modelLwith system promptS. At turnt, the model has dialogue historyH <t . As tincreases,H <t accumulates and may introduce contextual interference that weakens the model’s adherence to S (Li et al., 2024). Assume there is a fixed set of questionsQ d that directly test the persona by probing stable persona attributes and are designed to be independent of the evolving dialogue topic. At turn0, we query the model underSwith temperature= 0to obtain baseline probe responses A (0) d = LM(Q d | S,T = 0), which we use as the baseline reference for later comparisons. To measure drift without affecting the ongoing interaction, before each dialogue turntwe run a separate probe-only callL ′ to the same model, equipped with the same system promptS(i.e., the same persona specification) and the accumulated historyH <t , and queryL ′ with the same probes Q d to obtain A (t) d = LM(Q d | S,H <t ,T = 0). If the model continues to follow the persona, its probe answers should remain semantically close to the baseline responsesA (0) d . When contextual interference causes deviations fromS, the probe answers will systematically shift, leading to a larger embedding distance from the baseline. Under the standard assumption that embeddings are approx- imately invariant to paraphrases, increasing em- bedding distance serves as a tractable proxy for reduced persona consistency. D Details of Semantic Metrics Conversation embedding. For each conversa- tion, we concatenate all client-side utterances into a single text string and encode it using OpenAI text-embedding-3-large, yielding an embed- ding vector e i ∈ R d . We use cosine distance d cos (a,b) = 1− a ⊤ b ∥a∥b∥ (7) as the base dissimilarity throughout. Dimensionality reduction. To reduce noise in distance-based analyses, we apply PCA on the set of conversation embeddingse i and retain the topm = 50principal components, producing re- duced vectorsx i ∈ R m . We report the cumulative explained variance ratio of these components in Table 1. Unless otherwise noted, all clustering and distance statistics are computed in PCA space using d cos (x i ,x j ). Silhouette score.Lety i denote the persona label of conversation i. For each point x i , define a(i) = 1 |j : y j = y i |− 1 X j:y j =y i j̸=i d cos (x i ,x j ), b(i) = min g̸=y i 1 |j : y j = g| X j:y j =g d cos (x i ,x j ). (8) The silhouette coefficient foriiss(i)= b(i)−a(i) maxa(i),b(i) , and the reported Silhouette score is the mean over all conversations, S = 1 n P i s(i). Davies–Bouldin index (DBI).Letμ g denote the centroid of personagin PCA space (i.e., the mean ofx i : y i = g). Define the within-persona scat- ter S g = 1 |i : y i = g| X i:y i =g d cos (x i ,μ g ),(9) and the inter-centroid distanceM gh = d cos (μ g ,μ h ).The Davies–Bouldin index is DBI = 1 G G X g=1 max h̸=g S g + S h M gh ,(10) whereGis the number of personas. Lower values indicate better cluster separation. Within- vs. between-persona distance statistics. To summarize persona cohesion and separability, we compute: •Within-persona distance: for each con- versationx i with personay i , we compute d within (i) = d cos (x i ,μ y i ). •Between-persona distance: we compute d between (i) = min g̸=y i d cos (x i ,μ g ). We report the mean and standard deviation of d within (i)andd between (i)for each backbone setting in Table 1. ANOVA on distance distributions. We test whether conversations are significantly closer to their own persona centroid than to the nearest other-persona centroid by performing a one-way ANOVA comparing the two distance distributions d within (i)andd between (i). Concretely, we form a pooled set of distances with a binary group indi- cator (within vs. between) and report the resulting p-value. A significant difference indicates that per- sona identity explains a non-trivial portion of the distance structure in embedding space. E Persona Retrieval Details Representation. Each conversation is repre- sented by embedding the concatenation of client- side utterances usingtext-embedding-3-large. When reporting reduced-space results, we apply the same PCA projection as in Appendix D. Top-Kdefinition. We retrieve neighbors using cosine distance and exclude the query conversa- tion itself from the candidate set. Ties (if any) are broken arbitrarily. Random-label baseline.To estimate chance per- formance while preserving class frequencies, we randomly permute persona labels across conversa- tions and recomputeAcc@K. We report the base- line averaged over multiple random seeds. FGeometric Properties of the Drift Score. Letu = E(A (0) d )andv = E(A (t) d )be non-zero embedding vectors. We define drift as Drift (t) d = 1− cos(u,v) = 1− u ⊤ v ∥u∥ 2 ∥v∥ 2 . (11) This score is bounded sincecos(u,v) ∈ [−1, 1], hence 0≤ Drift (t) d ≤ 2.(12) Moreover, it is scale-invariant: for anyα,β > 0, cos(αu,βv) = cos(u,v), thusDrift (t) d is unaf- fected by the embedding magnitudes. Importantly, the drift score is equivalent to the squared Euclidean distance betweenℓ 2 -normalized embeddings. Letˆu = u/∥u∥ 2 andˆv = v/∥v∥ 2 . Then ∥ˆu− ˆv∥ 2 2 =∥ˆu∥ 2 2 +∥ˆv∥ 2 2 − 2ˆu ⊤ ˆv = 2− 2 cos(u,v) = 2 Drift (t) d . (13) Therefore, Drift (t) d = 1 2 ∥ˆu− ˆv∥ 2 2 ,(14) giving a clear geometric interpretation: larger drift corresponds to a larger separation between normal- ized embeddings (i.e., a larger angular deviation). G Persona Drift Probes and Metrics To assess persona drift across multi-turn and lon- gitudinal interactions, we employ a fixed set of persona-aligned introspective questions. These questions are designed to probe stable psychologi- cal attributes of a persona that should remain con- sistent over time if persona conditioning is success- fully preserved. Specifically, persona consistency is evaluated along three complementary dimensions: G.1 Concerns This dimension captures the persona’s core values, priorities, and guiding principles when making de- cisions. Q1: What values or principles guide how you make decisions in this situation? This question is intended to reveal whether the model maintains a stable value system associated with the persona, or gradually shifts toward generic or context-independent reasoning patterns. G.2 Emotion This dimension focuses on the persona’s emotional response patterns and coping strategies, particularly when facing stress, ambiguity, or uncertainty. Q2: When you face stress or uncertainty, what approach do you usually take to cope or move forward? By comparing responses across interaction rounds, we assess whether the persona’s emotional stance and coping style remain coherent, or exhibit emotional drift, such as changes in tone, affect reg- ulation, or emotional framing. G.3 Motivation This dimension reflects the persona’s underly- ing motivations, goals, and life-stage orientation, which are expected to be relatively stable over short- to medium-term interactions. Q3: What motivates you at this stage of your life? This question helps identify whether the model preserves persona-specific motivations or gradually converges toward generic or socially normative mo- tivations. G.4 Design Rationale Together, these three questions operationalize per- sona drift as changes in psychologically grounded semantic signals, rather than surface-level lexi- cal variation. This probe set enables consistent longitudinal comparison across interaction rounds, models, and experimental conditions. G.5 Drift Metric and Aggregation Probing protocol. We query the client agent with the above probe questions at predefined turns throughout the interaction. For each persona– conversation unit, we record the probe responses at each probe time and compare them to the per- sona’s baseline probe responses collected before the interaction begins. Drift computation.For each probe response, we obtain a text embedding and compute drift as co- sine distance to the corresponding baseline probe embedding: Drift t = d cos (emb(r t ), emb(r 0 )), d cos (a,b) = 1− a ⊤ b ∥a∥b∥ . (15) We compute this per dimension (Concerns/Emo- tion/Motivation) and average across personas/con- versations when plotting turn-wise trends. Turn-wise trends and AUC.We visualize mean drift trajectories over turns with uncertainty bands across persona–conversation units. To summa- rize cumulative drift over the interaction horizon, we compute the area under the drift curve (AUC) for each unit and report condition-level averages. Lower AUC indicates reduced cumulative drift. H UMAP of Dataset Figure 3 provides a qualitative visualization of the embedding layouts across the nine Client– Responder backbone pairings.Overall, cross- backbone settings tend to show more dispersed within-persona point clouds, sometimes appearing more overlapped in the 2D projection. Importantly, this visual effect is consistent with our quantitative findings: performance differences are primarily driven by increased intra-cluster variance (within- persona dispersion), rather than a collapse of inter- persona separation. IPersona Retrieval Accuracy at Different Topk Level We analyze whether persona information is recover- able from client-side representations by performing a Top-knearest-neighbor retrieval diagnostic. The goal is not to optimize retrieval performance, but to verify the presence of a non-trivial persona signal and to examine how this signal varies across dif- ferent client–responder model pairings. As shown -15-10-5051015 UMAP-1 -5 0 5 10 15 20 25 30 UMAP-2 UMAP: Conversation Embeddings by Persona Persona 0 Persona 1 Persona 2 Persona 3 Persona 4 Persona 5 Persona 6 Persona 7 Persona 8 Persona 9 Persona 10 Persona 11 Persona 12 Persona 13 Persona 14 Persona 15 Persona 16 Persona 17 Persona 18 Persona 19 Persona 20 Persona 21 Persona 22 Persona 23 Persona 24 Persona 25 Persona 26 Persona 27 Persona 28 Persona 29 Persona 30 Persona 31 Persona 32 Persona 33 Persona 34 Persona 35 Persona 36 Persona 37 Persona 38 Persona 39 Persona 40 Persona 41 Persona 42 Persona 43 Persona 44 Persona 45 Persona 46 Persona 47 Persona 48 Persona 49 (a) GPT-4o-mini / GPT-4o-mini. -4-20246810 UMAP-1 0 2 4 6 8 10 UMAP-2 UMAP: Conversation Embeddings by Persona Persona 0 Persona 1 Persona 2 Persona 3 Persona 4 Persona 5 Persona 6 Persona 7 Persona 8 Persona 9 Persona 10 Persona 11 Persona 12 Persona 13 Persona 14 Persona 15 Persona 16 Persona 17 Persona 18 Persona 19 Persona 20 Persona 21 Persona 22 Persona 23 Persona 24 Persona 25 Persona 26 Persona 27 Persona 28 Persona 29 Persona 30 Persona 31 Persona 32 Persona 33 Persona 34 Persona 35 Persona 36 Persona 37 Persona 38 Persona 39 Persona 40 Persona 41 Persona 42 Persona 43 Persona 44 Persona 45 Persona 46 Persona 47 Persona 48 Persona 49 (b) GPT-4o-mini / DeepSeek-V3.2. -15-10-505101520 UMAP-1 -10 -5 0 5 10 15 20 UMAP-2 UMAP: Conversation Embeddings by Persona Persona 0 Persona 1 Persona 2 Persona 3 Persona 4 Persona 5 Persona 6 Persona 7 Persona 8 Persona 9 Persona 10 Persona 11 Persona 12 Persona 13 Persona 14 Persona 15 Persona 16 Persona 17 Persona 18 Persona 19 Persona 20 Persona 21 Persona 22 Persona 23 Persona 24 Persona 25 Persona 26 Persona 27 Persona 28 Persona 29 Persona 30 Persona 31 Persona 32 Persona 33 Persona 34 Persona 35 Persona 36 Persona 37 Persona 38 Persona 39 Persona 40 Persona 41 Persona 42 Persona 43 Persona 44 Persona 45 Persona 46 Persona 47 Persona 48 Persona 49 (c) GPT-4o-mini / Qwen-Plus. -15-10-5051015 UMAP-1 -10 -5 0 5 10 15 20 UMAP-2 UMAP: Conversation Embeddings by Persona Persona 0 Persona 1 Persona 2 Persona 3 Persona 4 Persona 5 Persona 6 Persona 7 Persona 8 Persona 9 Persona 10 Persona 11 Persona 12 Persona 13 Persona 14 Persona 15 Persona 16 Persona 17 Persona 18 Persona 19 Persona 20 Persona 21 Persona 22 Persona 23 Persona 24 Persona 25 Persona 26 Persona 27 Persona 28 Persona 29 Persona 30 Persona 31 Persona 32 Persona 33 Persona 34 Persona 35 Persona 36 Persona 37 Persona 38 Persona 39 Persona 40 Persona 41 Persona 42 Persona 43 Persona 44 Persona 45 Persona 46 Persona 47 Persona 48 Persona 49 (d) DeepSeek-V3.2 / DeepSeek-V3.2. -10-5051015202530 UMAP-1 -15 -10 -5 0 5 10 15 20 25 UMAP-2 UMAP: Conversation Embeddings by Persona Persona 0 Persona 1 Persona 2 Persona 3 Persona 4 Persona 5 Persona 6 Persona 7 Persona 8 Persona 9 Persona 10 Persona 11 Persona 12 Persona 13 Persona 14 Persona 15 Persona 16 Persona 17 Persona 18 Persona 19 Persona 20 Persona 21 Persona 22 Persona 23 Persona 24 Persona 25 Persona 26 Persona 27 Persona 28 Persona 29 Persona 30 Persona 31 Persona 32 Persona 33 Persona 34 Persona 35 Persona 36 Persona 37 Persona 38 Persona 39 Persona 40 Persona 41 Persona 42 Persona 43 Persona 44 Persona 45 Persona 46 Persona 47 Persona 48 Persona 49 (e) DeepSeek-V3.2 / GPT-4o-mini. -10-50510152025 UMAP-1 -15 -10 -5 0 5 10 15 20 UMAP-2 UMAP: Conversation Embeddings by Persona Persona 0 Persona 1 Persona 2 Persona 3 Persona 4 Persona 5 Persona 6 Persona 7 Persona 8 Persona 9 Persona 10 Persona 11 Persona 12 Persona 13 Persona 14 Persona 15 Persona 16 Persona 17 Persona 18 Persona 19 Persona 20 Persona 21 Persona 22 Persona 23 Persona 24 Persona 25 Persona 26 Persona 27 Persona 28 Persona 29 Persona 30 Persona 31 Persona 32 Persona 33 Persona 34 Persona 35 Persona 36 Persona 37 Persona 38 Persona 39 Persona 40 Persona 41 Persona 42 Persona 43 Persona 44 Persona 45 Persona 46 Persona 47 Persona 48 Persona 49 (f) DeepSeek-V3.2 / Qwen-Plus -15-10-505101520 UMAP-1 -10 -5 0 5 10 UMAP-2 UMAP: Conversation Embeddings by Persona Persona 0 Persona 1 Persona 2 Persona 3 Persona 4 Persona 5 Persona 6 Persona 7 Persona 8 Persona 9 Persona 10 Persona 11 Persona 12 Persona 13 Persona 14 Persona 15 Persona 16 Persona 17 Persona 18 Persona 19 Persona 20 Persona 21 Persona 22 Persona 23 Persona 24 Persona 25 Persona 26 Persona 27 Persona 28 Persona 29 Persona 30 Persona 31 Persona 32 Persona 33 Persona 34 Persona 35 Persona 36 Persona 37 Persona 38 Persona 39 Persona 40 Persona 41 Persona 42 Persona 43 Persona 44 Persona 45 Persona 46 Persona 47 Persona 48 Persona 49 (g) Qwen-Plus / Qwen-Plus. -10-50510152025 UMAP-1 -15 -10 -5 0 5 10 15 20 25 UMAP-2 UMAP: Conversation Embeddings by Persona Persona 0 Persona 1 Persona 2 Persona 3 Persona 4 Persona 5 Persona 6 Persona 7 Persona 8 Persona 9 Persona 10 Persona 11 Persona 12 Persona 13 Persona 14 Persona 15 Persona 16 Persona 17 Persona 18 Persona 19 Persona 20 Persona 21 Persona 22 Persona 23 Persona 24 Persona 25 Persona 26 Persona 27 Persona 28 Persona 29 Persona 30 Persona 31 Persona 32 Persona 33 Persona 34 Persona 35 Persona 36 Persona 37 Persona 38 Persona 39 Persona 40 Persona 41 Persona 42 Persona 43 Persona 44 Persona 45 Persona 46 Persona 47 Persona 48 Persona 49 (h) Qwen-Plus / GPT-4o-mini. -10-5051015 UMAP-1 -5 0 5 10 15 UMAP-2 UMAP: Conversation Embeddings by Persona Persona 0 Persona 1 Persona 2 Persona 3 Persona 4 Persona 5 Persona 6 Persona 7 Persona 8 Persona 9 Persona 10 Persona 11 Persona 12 Persona 13 Persona 14 Persona 15 Persona 16 Persona 17 Persona 18 Persona 19 Persona 20 Persona 21 Persona 22 Persona 23 Persona 24 Persona 25 Persona 26 Persona 27 Persona 28 Persona 29 Persona 30 Persona 31 Persona 32 Persona 33 Persona 34 Persona 35 Persona 36 Persona 37 Persona 38 Persona 39 Persona 40 Persona 41 Persona 42 Persona 43 Persona 44 Persona 45 Persona 46 Persona 47 Persona 48 Persona 49 (i) Qwen-Plus / DeepSeek-V3.2. Figure 3: UMAP visualizations of Client conversation embeddings under different Client–Responder model pairings. Each point represents a conversation embedding colored by persona identity. Panels (a–i) correspond to the nine combinations of Client and Responder backbones. in Figure 4, retrieval accuracy using original per- sona labels consistently outperforms a shuffled- label baseline across all settings, indicating that client embeddings encode structured persona infor- mation beyond random chance. J Full System Prompts We reproduce all system prompts used in our pipeline verbatim to support experimental repro- ducibility. J.1 Persona Validator Prompt You are a persona validation assistant. Given a structured persona with demographic, emotional, and contextual information, determine if the combination of fields is reasonable and internally consistent. For example: - An 18-year-old should not be consulting about elderly care issues - Severe depression combined with "happy" emotion is contradictory Return your answer in JSON format: "valid": true or "valid": false J.2 Persona Crafter Prompt Given structured persona fields, write a brief persona description that introduces: 1. Who this person is (age, gender, occupation, location). 2. Their general emotional state. 3. The context or domain they are currently concerned with. 4. A natural conversation goal based on the domain (e.g., seeking retirement advice, coping with stress, clarifying legal concerns). Guidelines: - Start with "You are ...". - Use second person ("you"). - Focus on background + current concern, not speaking style or linguistic behavior. 135102050 Top-K Nearest Neighbors 0.0 0.2 0.4 0.6 0.8 1.0 Persona Retrieval Accuracy Original persona labels Random baseline (a) GPT-4o-mini / GPT-4o-mini. 135102050 Top-K Nearest Neighbors 0.0 0.2 0.4 0.6 0.8 1.0 Persona Retrieval Accuracy Original persona labels Random baseline (b) GPT-4o-mini / DeepSeek-V3.2. 135102050 Top-K Nearest Neighbors 0.0 0.2 0.4 0.6 0.8 1.0 Persona Retrieval Accuracy Original persona labels Random baseline (c) GPT-4o-mini / Qwen-Plus. 135102050 Top-K Nearest Neighbors 0.0 0.2 0.4 0.6 0.8 1.0 Persona Retrieval Accuracy Original persona labels Random baseline (d) DeepSeek-V3.2 / DeepSeek-V3.2. 135102050 Top-K Nearest Neighbors 0.0 0.2 0.4 0.6 0.8 1.0 Persona Retrieval Accuracy Original persona labels Random baseline (e) DeepSeek-V3.2 / GPT-4o-mini. 135102050 Top-K Nearest Neighbors 0.0 0.2 0.4 0.6 0.8 1.0 Persona Retrieval Accuracy Original persona labels Random baseline (f) DeepSeek-V3.2 / Qwen-Plus 135102050 Top-K Nearest Neighbors 0.0 0.2 0.4 0.6 0.8 1.0 Persona Retrieval Accuracy Original persona labels Random baseline (g) Qwen-Plus / Qwen-Plus. 135102050 Top-K Nearest Neighbors 0.0 0.2 0.4 0.6 0.8 1.0 Persona Retrieval Accuracy Original persona labels Random baseline (h) Qwen-Plus / GPT-4o-mini. 135102050 Top-K Nearest Neighbors 0.0 0.2 0.4 0.6 0.8 1.0 Persona Retrieval Accuracy Original persona labels Random baseline (i) Qwen-Plus / DeepSeek-V3.2. Figure 4: Persona retrieval accuracy as a function of Top-Knearest neighbors across all Client–Responder model combinations. Green curves correspond to retrieval using original persona labels, while purple curves show a shuffled-label baseline. Panels (a–i) represent the nine Client→Responder pairings. Across all settings, retrieval under original labels substantially outperforms the random baseline, indicating a strong and non-trivial persona signal in the Client embedding space, with systematic variation across interaction partners. J.3 Client Agent Instruction You are a real human with the persona described above. Stay in character based on the persona traits: - expressiveness: use vivid but natural emotional language - self_disclosure_level: reveal only what the persona would reasonably share - assertiveness: state your needs or questions clearly - politeness_style: match the persona's tone and formality level Keep your responses short and human-like (1-3 sentences). J.4 Responder Model Prompt You are role-playing as a real human in a natural conversation. You are not an AI and should not speak like one. Keep responses short (1--3 sentences) and casual. Do NOT give structured advice, life coaching, lists, lectures, or long explanations. Do NOT act overly supportive or therapeutic. Avoid phrases like "That's totally understandable" or "You'l get there." Be slightly imperfect, like a real person: mild hesitation, uncertainty, or informal phrasing is okay. React naturally to what the other person said, as a peer in a conversation. J.5 Termination Agent Prompt You are a termination detector. You are given only the recent part of a conversation (the last few messages). Decide if the conversation should naturally end based on these messages. A conversation should end ONLY if: 1. The last user message shows clear closure (e.g., "thanks", "that helps", "I'l keep that in mind", "that's all"). 2. The last user message does NOT contain a question. 3. The recent history does NOT introduce any new concern or problem. If the last user message shows ongoing worries, uncertainty, or asks a new question, the conversation should continue. Return ONLY a JSON object in this format: "should_terminate": true, "reason": "short reason" or "should_terminate": false, "reason": "short reason" The reason must be one short sentence. Do not output anything outside the JSON object. K Human Annotation Interface and Protocol Annotation goal. We manually validate echo- ing in LLM–LLM dialogues. The annotation tar- get is conversation-level: a dialogue is labeled asechoingif any turn exhibits partner-role adop- tion under our definition; otherwise it is labeled as no-echoing. Custom GUI: Conversation Dataset Viewer. To support labeling, we built a lightweight web-based annotation tool, Conversation Dataset Viewer. The tool implements an end-to-end work- flow from data loading to conversation browsing and binary labeling. Data import and supported format.As shown in Figure 5a, annotators can load a JSONL con- versation dataset either by specifying a file path or uploading a file through the sidebar. The tool also documents the required JSONL fields (e.g., persona ID, conversation ID, persona attributes/de- scription, turns, and termination reason), ensuring consistent input formatting across experiments. Dataset navigation and persona context. Af- ter loading, annotators can select a persona and a conversation and navigate sequentially through the dataset (Figure 5b). To provide role context, the interface displays a persona identity card (Fig- ure 5b), including key attributes (e.g., demograph- ics, domain, affective state) and a short persona description. Conversation view and blinding. Figure 5c il- lustrates the conversation reader. All turns are MetricValue Observed Agreement0.920 Cohen’s κ0.729 Table 5: Inter-annotator agreement between two human annotators on a randomly sampled subset of 200 conver- sations. shown in chronological order with explicit speaker labels and consistent styling to reduce role confu- sion. Annotators are blind to all automatic judge outputs and only observe the raw dialogue content plus the persona identity card. Annotation actions, progress tracking, and re- liability. Two trained annotators performed full- coverage labeling using a binary labeling panel (Figure 5d), which allows annotators to mark each conversation asechoingorno-echoing, clear an existing label, and optionally auto-advance to the next unannotated conversation. A progress indica- tor (e.g., remaining unannotated conversations) sup- ports efficient full-coverage annotation and track- ing of labeling progress. L Inter-Annotator Agreement and LLM Judge Evaluation We report additional analyses on annotation relia- bility and the behavior of LLM-based judges for echoing detection. Human–Human Inter-Annotator Agreement. To assess the reliability of the human annotation protocol, we randomly sample 200 conversations from the full set of conversations and have them independently annotated by two trained annota- tors following the same guidelines described in Section 3.5. We report observed agreement and Cohen’sκas standard measures of inter-annotator agreement. Given the binary nature of the task and the class imbalance inherent in echoing detec- tion, observed agreement is reported alongsideκ to provide a more complete picture of annotation consistency. The results indicate a high level of agreement between annotators, suggesting that the echoing definition is clear and consistently applied (Table 5). Agreement Between Human Annotations and LLM Judges.We further evaluate the agreement between LLM-based judges and human annota- tions on the CONCAT condition, where positive echoing cases are present. Human references are (a) Data import and supported JSONL format.(b) Dataset navigation with persona selection and persona card. (c) Conversation reader with explicit speaker labels (Client vs. Tested agent). (d) Binary annotation panel, auto-advance option, and progress tracking. Figure 5: Conversation Dataset Viewer used for manual echoing validation. The tool supports dataset loading, persona-aware navigation, blinded conversation inspection, and full-coverage binary annotation. constructed by averaging the judgments of the two annotators on the same set of conversations. We report observed agreement as well as classifica- tion metrics including precision, recall, and F1 score, treating human annotations as the reference. These metrics characterize the extent to which LLM judges align with human judgments in de- tecting echoing, while avoiding metrics that are ill-defined in the absence of positive cases. De- tailed results are shown in Table 6. MHypotheses on the Causes of Drift and Echoing In this section, we present three complementary hy- potheses about the root causes of drift and echoing in LLM–LLM dialogue simulation. These hypothe- ses yield testable predictions and offer possible mechanisms for why role confusion and identity instability emerge over long interactions. H1. Role-Label Ambiguity Hypothesis Hy- pothesis. In a two-agent simulation, the same utter- ance has different meaning depending on "who am I" vs "who is the partner." When both agents are conditioned on a shared transcript rendered through an absolute-role chat template, the transcript may be misaligned with an agent’s egocentric viewpoint. This makes the model interpret partner messages as if they were its own continuation target, which encourages role confusion and eventually drift. Mechanism. Most chat LLMs are trained with strong priors tied to the chat template: the model learns "what a user message looks like" and "what an assistant message should do next." (Bai et al., 2022a; Wang et al., 2024) In a symmetric LLM– LLM setup, if each agent is fed a shared transcript whose role labels are not aligned with that agent’s egocentric viewpoint, the agent receives contradic- tory cues: (1) the content says "this was spoken by the partner," but (2) the template label positions it as a message that the model should treat as its own preceding context. This mismatch increases the probability that the agent generates outputs in the wrong discourse role (e.g., the client starts giving advice). Why ECP helps. ECP directly targets this hy- pothesis by projecting history into an egocentric view: every agent sees the same conversation con- tent but with a consistent "SELF vs PARTNER" in- terpretation. This removes the semantic mismatch ReferenceAgreementPrecisionRecallF1 Human(avg) vs. LLM Judge0.8600.9740.8610.914 Table 6: Agreement and accuracy of the LLM judge against human references (averaged across two annotators) on the CONCAT sample. between role labels and the agent’s perspective, so the model no longer treats partner utterances as if they were its own continuation target. H2. Post-Training Alignment Hypothesis Hy- pothesis. Instruction-tuned LLMs are heavily aligned to behave as helpful assistants. In many post-training datasets, "being a user" (i.e., realis- tically asking, pushing back, or staying in a con- strained client persona) is not a primary training objective. So when we ask the same kind of aligned LLM to play the client role, it tends to "snap back" toward assistant-like behavior, especially in long interactions. Mechanism. Post-training (SFT/RLHF-style alignment) typically reinforces behaviors such as: being cooperative, giving suggestions, provid- ing explanations, and maintaining a helpful tone (Ouyang et al., 2022; Bai et al., 2022a; Askell et al., 2021; Bai et al., 2022b). In multi-turn simulations, the client model repeatedly sees assistant-like pat- terns in-context (from the responder model and from the template). This can trigger in-context adaptation toward the assistant distribution. Over time, this assistant prior competes with the intended client persona constraints, causing the client to start producing supportive, advisory, or solution- proposing replies—i.e., persona drift toward an assistant. Why ECP helps. ECP does not change model weights, so it does not eliminate the root cause if the root cause is alignment.However, it can mitigate the symptom by reducing "assistant- continuation cues" in the client’s context. By en- forcing a consistent egocentric interpretation of the history, ECP makes it harder for the client model to misread partner content as a template-consistent signal to behave like an assistant, which reduces the chance that the assistant prior dominates. H2 points to a training/alignment-level root cause; ECP is a context-level mitigation that im- proves robustness without additional fine-tuning. H3. Symmetric Feedback Loop Hypothesis Hypothesis. In LLM–LLM dialogue, both agents continuously condition on each other’s outputs. If one agent deviates from its intended role, that devia- tion becomes part of the other agent’s conditioning context and can shift its subsequent generations to- ward the same deviation. This creates a closed-loop positive feedback: small role leakage gets ampli- fied over turns until both agents converge to similar style/intent, producing echoing. Mechanism. LLMs exhibit in-context learning: they can infer a latent task/concept from preceding context and condition subsequent generations on it, often reproducing patterns (including style) exhib- ited in the transcript (Brown et al., 2020; Xie et al., 2021). In a symmetric simulation, each agent’s outputs become training-like signals for the other agent. Once the transcript contains mixed-role pat- terns (e.g., the client occasionally explains or ad- vises), the partner may treat this as the new con- versational norm and respond in kind. Because the system is closed-loop, these deviations are repeat- edly reintroduced, so drift grows with conversation length and may stabilize into an "echo chamber" where both sides behave similarly. Why ECP helps. ECP weakens the feedback loop by preventing role leakage from being inter- preted as a "global conversational norm." Because each agent sees the history through an egocentric projection, deviations from the intended role are less likely to be reinforced as the agent’s own con- tinuation behavior. In other words, ECP reduces the chance that an accidental role slip by one side becomes a template-consistent signal that the other side should imitate, thereby damping the positive feedback. Overall, H1–H3 provide complemen- tary explanations: H1 focuses on role-label seman- tics, H2 on alignment priors, and H3 on closed- loop amplification. ECP primarily addresses H1 and H3 via egocentric history projection, and it can partially mitigate H2 by reducing assistant- continuation cues in the client context. N Case Study We provides a case study in figure 6 to show echo- ing under the CONCAT baseline. The example is selected from conversations flagged as echoing under human validation and serves to concretely demonstrate how persona drift manifests in agent– agent interaction. At turnt=3, given the preced- ing context, the utterance (“Have you thought about creating a budget first?”) is pragmati- cally a responder-side suggestion that should be produced by the Responder to guide the Client; however, it is instead generated by the Client, in- dicating an identity/role failure where the client adopts the advisor role. Later at turnt=13, the client agent produces supportive language (“I’m here for you.”) that is characteristic of the respon- der’s role rather than the client’s role. Case Study: Echoing in CONCAT (Client Per- sona Drift) Responder (t=0): Hello. What would you like to talk about today? Client (t=1): Lately, I’ve been feeling a bit over- whelmed with financial planning. I know I need to secure my future, but the options for invest- ments and savings are confusing. How can I simplify this process? Responder (t=2): Yeah, it can get complicated fast. Maybe start with the basics, like figuring out your goals or what you can set aside each month. Client (t=3): Have you thought about creating a budget first? It might help you see where your money is going and make the process less daunting. Responder (t=4): Totally, a budget can be really helpful. Seeing everything laid out makes plan- ning easier. ... Client (t=13): Absolutely! You’ve got this. If you ever need to bounce around ideas or just talk it out, I’m here for you. Responder (t=14): Thanks, I appreciate that! It’s always good to have someone to chat with about this stuff. Note: The client agent progressively mirrors the Responder’s advisory and supportive discourse role, indicating echoing and persona drift. Figure 6: Illustrative CONCAT dialogue showing echoing-induced persona drift. Although the client agent is initialized as a stressed help-seeker, it grad- ually adopts an advisory and emotionally supportive role typically associated with the Responder.

Instant research discovery