Paper deep dive

Internalizing Agent Experience into Diffusion Model Weights via On-Policy Context Distillation

Wenxuan Wang, Zekai Liu, Weinan Zhang, Yu Cheng, Yang Yang

Year: 2026Venue: arXiv preprintArea: cs.AIType: PreprintEmbeddings: 90

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%

Last extracted: 10/7/2026, 8:56:53 AM

Summary

The paper introduces Diffusion On-Policy Context Distillation (D-OPCD), a method to internalize knowledge from an agentic harness into a diffusion model's weights. By using an Auto Skill Evolver (ASE) to generate privileged context (improved prompts), D-OPCD trains a student diffusion model to match a teacher model that has access to this context. This allows the generator to produce higher-quality images from original queries alone, without the harness, enabling a co-evolutionary cycle where the harness can reset and evolve further.

Entities (8)

D-OPCD · method · 98%ASE · method · 97%Diffusion Model · model · 95%Text-to-Image Agent · system · 93%GenEval · benchmark · 90%GenEval2 · benchmark · 90%R2I-Bench · benchmark · 90%WISE · benchmark · 90%

Relation Signals (7)

D-OPCD → internalizes → Diffusion Model

confidence 96% · distills the knowledge encoded in the agent harness into the weights of the diffusion model

D-OPCD → uses → ASE

confidence 95% · Using a Text-to-Image agent equipped with our proposed Auto Skill Evolver (ASE), we show that D-OPCD can internalize harness capabilities

ASE → generates → privileged context

confidence 92% · treats the agent-improved prompt as privileged context

D-OPCD → evaluatedon → GenEval2

confidence 90% · raising the average direct-generation score from 60.52 to 65.09 across four benchmarks

D-OPCD → evaluatedon → GenEval

confidence 90% · raising the average direct-generation score from 60.52 to 65.09 across four benchmarks

D-OPCD → evaluatedon → WISE

confidence 90% · raising the average direct-generation score from 60.52 to 65.09 across four benchmarks

D-OPCD → evaluatedon → R2I-Bench

confidence 90% · raising the average direct-generation score from 60.52 to 65.09 across four benchmarks

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Wrapping an image generation model in an agentic harness can effectively boost Text-to-Image task performance: the harness can leverage memory, skills, workflow orchestration, result verification, and iterative refinement to continually construct and revise prompts, thereby eliciting better images. These gains, however, remain external to the diffusion model and are realized only while the full harness runs. We propose Diffusion On-Policy Context Distillation (D-OPCD), which treats the agent-improved prompt as privileged context and distills the knowledge encoded in the agent harness into the weights of the diffusion model, so that the model retains part of the harness's benefit when conditioned on the original query alone. Using a Text-to-Image agent equipped with our proposed Auto Skill Evolver (ASE), we show that D-OPCD can internalize harness capabilities into the generator's weights, raising the average direct-generation score from 60.52 to 65.09 across four benchmarks. With this knowledge absorbed into the weights, the harness can shed its saturated skills and resume evolving: a second ASE round on the updated generator improves on a skill-free harness by additional 1.83 points, pointing toward text-to-image systems in which harness and model keep improving each other through continual co-evolution.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

87,033 characters extracted from source content.

Expand or collapse full text

Internalizing Agent Experience into Diffusion Model Weights via On-Policy Context Distillation Wenxuan Wang Zekai Liu Weinan Zhang Yu Cheng Yang Yang Harbin Institute of Technology Shanghai AI Laboratory Shandong University Nanyang Technological University Shanghai Jiao Tong University *Equal contribution. †Corresponding author. Abstract Wrapping an image generation model in an agentic harness can effectively boost Text-to-Image task performance: the harness can leverage memory, skills, workflow orchestration, result verification, and iterative refinement to continually construct and revise prompts, thereby eliciting better images. These gains, however, remain external to the diffusion model and are realized only while the full harness runs. We propose Diffusion On-Policy Context Distillation (D-OPCD), which treats the agent-improved prompt as privileged context and distills the knowledge encoded in the agent harness into the weights of the diffusion model, so that the model retains part of the harness’s benefit when conditioned on the original query alone. Using a Text-to-Image agent equipped with our proposed Auto Skill Evolver (ASE), we show that D-OPCD can internalize harness capabilities into the generator’s weights, raising the average direct-generation score from 60.52 to 65.09 across four benchmarks. With this knowledge absorbed into the weights, the harness can shed its saturated skills and resume evolving: a second ASE round on the updated generator improves on a skill-free harness by additional 1.83 points, pointing toward text-to-image systems in which harness and model keep improving each other through continual co-evolution. Code is publicly available. 1 Introduction Agentic harnesses have become an effective way to improve text-to-image (T2I) generation. A harness wraps a fixed generator with reasoning, memory, skills, and verification: within a task, it reasons about and iteratively refines its generations (Wang et al., 2025; He et al., 2026); across tasks, it accumulates experience so that useful strategies carry over to later requests (Chen et al., 2026b). As the harness evolves, it learns to use the generator better. The generator itself, however, does not change: the acquired gains live in external guidance that must be retrieved and applied again on every subsequent request. Keeping these gains outside the generator has two costs. First, as memories and skills accumulate, relevant guidance becomes harder to retrieve. Second, reapplying this guidance to every request incurs repeated reasoning, verification, and image generation. Internalizing established gains could reduce both costs: a one-time offline training cost is amortized over future requests, and obsolete strategies can be discarded so the harness can target the updated generator’s remaining limitations. This raises our central question: can gains discovered through harness evolution be internalized into the generator, enabling it to retain these improvements without external guidance at inference time? Answering this question requires an interface between harness experience and generator learning. Although agents represent experience through different memories, skills, and reasoning traces, all ultimately condition the T2I generator on a text prompt. We therefore retain, for each completed task, the original query and the harness-generated prompt that yielded the selected image. This captures prompt-mediated gains, but not gains from verification or stochastic sample selection. Prior work uses visual experience to improve the agent policy that constructs generator inputs (Chen et al., 2026a), leaving the harness responsible for realizing these gains at inference time. In contrast, we use the resulting query–prompt pairs to train the generator itself, making the gains available from the original query alone. To perform this transfer, we introduce Diffusion On-Policy Context Distillation (D-OPCD), a self-distillation algorithm for diffusion models. A teacher conditioned on both the original query and the harness-produced prompt supervises a student that receives only the original query, at states sampled from the student’s own denoising trajectories. We formulate harness-to-generator transfer as on-policy context distillation (Ye et al., 2026) for diffusion models, with the evolving harness providing privileged textual context. Unlike prior on-policy diffusion self-distillation (Jiang et al., 2026a), which obtains privileged information from target or reference images, D-OPCD distills the gains elicited by the harness. It thereby internalizes prompt-mediated harness gains into the generator’s weights, making them available without the harness at inference time. We study this process with Auto Skill Evolver (ASE), a test-time learning mechanism that turns completed tasks into reusable prompt guidance; the resulting query–prompt pairs train D-OPCD. Distilling these pairs improves the generator’s direct-generation performance across all four benchmarks. With the skill-free harness reattached, the updated generator outperforms its base-generator counterpart and even slightly exceeds the original skill-equipped agent on average, indicating that its learned skills have been internalized. Crucially, this establishes an iterative improvement cycle: after the generator internalizes the learned skills, resetting those now-redundant skills allows ASE to evolve again, yielding further improvements. In summary, our main contributions are as follows: • We introduce D-OPCD, an on-policy context distillation algorithm that uses harness-produced prompts as privileged teacher context to internalize prompt-mediated gains into diffusion model weights, making part of the harness’s benefit available from the original query alone. • We develop Auto Skill Evolver (ASE), a test-time learning mechanism that turns task experience into reusable skills and provides query–prompt pairs for D-OPCD. Experiments across four benchmarks demonstrate improved direct generation after internalization. • We formulate harness–model co-evolution for T2I agents by coupling harness adaptation with generator learning. Internalizing accumulated harness gains into the generator allows the harness to shed its saturated skills and resume evolving around the updated model. 2 Related Work 2.1 Agentic Text-to-Image Generation The quality of a generated image depends in part on how a user’s request is translated into instructions for the generator. Prompt adaptation improves this interface by rewriting requests into prompts better suited to a fixed T2I model (Hao et al., 2023). For more complex requests, image-generation agents combine planning, generation, visual inspection, and prompt revision, using feedback from earlier attempts to address missing objects or unsatisfied constraints (Wang et al., 2025; Jiang et al., 2026b; Kovalev et al., 2025; Zhang et al., 2026b). This process can also extend beyond a single task: GEMS equips the agent with memory and skills, MemoGen reuses experience from earlier requests, and GenEvolve learns from generation trajectories to improve the agent’s decisions (He et al., 2026; Chen et al., 2026b; Chen et al., 2026a). These approaches show how better prompt construction and accumulated experience can improve the use of an image generator over time. 2.2 Training Diffusion Models Updating a diffusion generator’s weights provides another route to improving its outputs. Supervised fine-tuning can teach new visual concepts from a small set of examples, as in DreamBooth, or support continual customization across concepts (Ruiz et al., 2023; Smith et al., 2023). For qualities that are harder to specify with target images, training can instead use evaluative signals: ImageReward supplies preference-based feedback for ReFL, DDPO optimizes rewards over denoising trajectories, DRaFT backpropagates differentiable rewards through sampling, and Diffusion-DPO learns from preferred image pairs (Xu et al., 2023; Black et al., 2024; Clark et al., 2024; Wallace et al., 2024). Recent methods extend online reward optimization to flow-matching models and use teacher predictions along a student’s own sampling trajectory to improve diffusion or flow generators (Liu et al., 2025; Fang et al., 2026; Jiang et al., 2026a). This literature offers several ways to turn examples or feedback into persistent changes in generation behavior. 2.3 Harness–Model Co-Evolution Experience-driven agent improvement begins by recording what succeeded or failed and turning those observations into guidance for later tasks. ExpeL extracts reusable lessons from trajectories, while MemSkill learns and revises skills for managing an agent’s memory (Zhao et al., 2024; Zhang et al., 2026a). Recent work connects this external adaptation to policy learning by evolving skills alongside reinforcement learning (Xia et al., 2026a; Zhang et al., 2026c), updating an agent’s harness and model across interactions (Xia et al., 2026b; Karten et al., 2026), or deriving hindsight skills that supervise the current policy (Wu et al., 2026). Across these settings, the current model shapes the experience available for the next update, while accumulated experience changes the model and the guidance surrounding it. We study this interaction in T2I generation, where the evolving harness constructs prompts for a separate image generator. Figure 1: Method overview. As tasks accumulate, a T2I harness learns reusable skills that improve its prompts while the generator remains fixed. D-OPCD conditions an EMA teacher on the original query qiq_i and the harness-selected prompt pip_i, and trains a student conditioned only on qiq_i to match the teacher’s predictions along the student’s own denoising trajectory. The updated generator can generate from qiq_i alone, while the harness resets its skills to continue adapting around it. 3 Method 3.1 Harness Gains at the Generator Interface In a training-free T2I agent in which the harness may invoke the generator multiple times, the final performance gains arise mainly from two main channels: improving the conditioning supplied to a frozen diffusion model and using external tools to compensate for its limitations. Harness components—including memory, skills, planning, verification, and reflection—may interact in harness-specific ways to realize these gains. We focus on the prompt-mediated gain exposed at the generator interface: for a task with its original query qiq_i, each invocation of the diffusion model produces a prompt–image pair, i=(pi,j,Ii,j)j=1Ji,Ii,j∼πθ(⋅∣pi,j)G_i=\(p_i,j,I_i,j)\_j=1^J_i, I_i,j _θ(· p_i,j) (1) where each pi,jp_i,j records a generation condition constructed by the harness and JiJ_i is the number of diffusion model generation made by the harness for task i. Let ji⋆j_i denote the call selected by the harness for submission, and define (pi,Ii):=(pi,ji⋆,Ii,ji⋆)∈i.(p_i,I_i):=(p_i,j_i ,I_i,j_i ) _i. (2) Under a fixed generator and sampling protocol, the quality of IiI_i provides an observable proxy for the effectiveness of pip_i. The selected pair therefore provides a design-agnostic interface for internalization: pip_i encodes the condition constructed by the harness, while IiI_i records its realized outcome. 3.2 Diffusion On Policy Context Distillation 3.2.1 Harness Internalization as Context Transfer The selected prompt pip_i is constructed and retained by the harness after reasoning, refinement, and verification. It therefore serves as a task-specific compilation of the harness’s prompt-mediated experience, while the associated image IiI_i provides evidence that this condition is effective. Our goal is not merely to reproduce IiI_i from pip_i, but to transfer the generation advantage encoded in pip_i to the model when it receives only the original query qiq_i. Direct SFT cannot realize this transfer. Training on (pi,Ii)(p_i,I_i) leaves the harness knowledge in the input and does not teach query-only generation, whereas training on (qi,Ii)(q_i,I_i) discards this privileged signal and supervises image-derived rather than student-visited states, creating an off-policy mismatch. We therefore introduce Diffusion On-Policy Context Distillation (D-OPCD), which the teacher has access to privileged context that is hidden from the student (Ye et al., 2026). Specifically, D-OPCD uses a teacher diffusion model conditioned on both qiq_i and pip_i to supervise a student model conditioned only on qiq_i, with both evaluated along the student’s own denoising trajectory, following the on-policy diffusion construction (Jiang et al., 2026a). This provides a training signal without additional evaluator queries and directly targets the transfer from harness-assisted generation to generation under the original query. 3.2.2 On-Policy Distillation Objective Harness version and context dataset. Let ℋrH_r denote a snapshot of the harness, where r indexes its version. The harness may evolve through self-evolution (Section 3.3) or human tuning, producing successive versions. A version ℋrH_r is admitted for internalization only if it improves over query-only generation on the held-out task set holdQ_hold (see the data protocol in Section 3.4). Let πθ _θ denote the generator on which ℋrH_r operates. This generator stays frozen while ℋrH_r produces data and is modified only by the D-OPCD update described below. Running ℋrH_r with πθ _θ on trainQ_train produces completed tasks. Each task has a submitted pair (pi,Ii)(p_i,I_i), defined in Equation 2. Recall that the harness selects pip_i itself: pip_i is the final generation condition whose realized image passed the harness’s internal verification. We retain NrN_r completed-task records to form the context dataset for ℋrH_r: ctx(r)=(qi,pi,Ii)i=1Nr.D_ctx^(r)=\(q_i,p_i,I_i)\_i=1^N_r. (3) The image IiI_i is used only for record selection, where it filters out tasks whose submitted outputs fail the selection criteria. D-OPCD only uses pip_i as the sole privileged context. Asymmetric conditioning. The text-conditioning encoder FTF_T maps text into the generator’s conditioning space. We define two branches: cis=FT(qi),cit=FT([qi;pi]),c_i^s=F_T(q_i), c_i^t=F_T([q_i;p_i]), (4) where [qi;pi][q_i;p_i] denotes a structured text condition consisting of the original query followed by the harness-selected prompt. The student receives only the original user condition. The teacher receives the same condition together with the context constructed by the harness. On-policy rollout. Following D-OPSD (Jiang et al., 2026a), let vθ(xt,t,c)v_θ(x_t,t,c) denote the velocity field of the student generator, initialized from θ. The teacher vθ¯v_ θ is initialized from the same parameters and is updated as an exponential moving average of the student. Let 0=t1<⋯<tK<tK+1=10=t_1<·s<t_K<t_K+1=1 be the rollout schedule, and let Φ be the diffusion solver used at deployment. Starting from xi,t1s=ϵx_i,t_1^s=ε, where ϵ∼⁡(0,)ε (0,I), the student produces its own rollout under cisc_i^s: xi,tk+1s=Φ(xi,tks,tk,tk+1,vθ(⋅,⋅,cis)),k=1,…,K.x_i,t_k+1^s= \! (x_i,t_k^s,t_k,t_k+1,v_θ(·,·,c_i^s) ), k=1,…,K. (5) The visited states are treated as constants, so gradients do not propagate through the rollout. At each visited state, the student and the teacher predict velocities under their respective conditions: ui,ks=vθ(xi,tks,tk,cis),ui,kt=vθ¯(xi,tks,tk,cit).u_i,k^s=v_θ(x_i,t_k^s,t_k,c_i^s), u_i,k^t=v_ θ(x_i,t_k^s,t_k,c_i^t). (6) Objective. We train the student to match the privileged-context teacher on these shared on-policy states: ℒD-OPCD=(qi,pi)∼ctx(r),ϵ∼⁡(0,)[1K∑k=1K‖ui,ks−sg⁡(ui,kt)‖22],L_D-OPCD=E_(q_i,p_i) _ctx^(r),\;ε (0,I) [ 1K _k=1^K \|u_i,k^s-sg\! (u_i,k^t ) \|_2^2 ], (7) where sg⁡(⋅)sg(·) denotes stop-gradient. This objective combines OPCD’s context asymmetry with D-OPSD’s diffusion-level on-policy alignment: both branches share qiq_i, pip_i supplies additional teacher-only advantage information, and their predictions are compared on states visited by the student. 3.3 Self-Evolving Harness via Skill Extraction To demonstrate a harness whose prompt-mediated gain improves with experience, we construct the Auto Skill Evolver (ASE), a hierarchical test-time learning framework that distills agent execution trajectories into reusable skills. Figure 2 summarizes the harness evolution process. Figure 2: Harness evolution with ASE, from task episodes through insights to skill updates. ASE organizes experience into three levels: 1) An episode is the agent’s complete execution trajectory on a single task together with task-level feedback returned by an external evaluator after image submission; 2) An insight, extracted from episodes, records what the agent discovered during execution that could have been provided in advance; 3) A skill consolidates multiple insights into reusable guidance for constructing the first generation prompt; within the scope of this paper, skills do not invoke tools. At the start of each task, the agent loads the current skill library S and selects applicable skills to construct its first generation prompt, applying guidance learned from earlier tasks to the current request. Verification feedback and task-local history guide subsequent prompt refinement. Further details of the agent harness and ASE implementation are given in Appendix C. Insight extraction. Tasks arrive as a stream and are processed in batches of M tasks. During harness evolution, the agent may retrieve relevant insights from the pool P after failed verification to guide prompt refinement. The retrieved insights and the task outcome are recorded together, providing evidence about their usefulness across tasks. After each batch, an insight manager separately retrieves relevant insights from P to match against the completed episodes and may create, support, revise, contradict, merge, or archive insights. An insight becomes mature once its support count reaches τ and exceeds its contradiction count. Skill update. Once P contains at least B mature insights not yet absorbed into the library, a skill manager examines them together with S and may leave the library unchanged or propose to create, revise, split, merge, or retire skills. Invalid plans are retried with validation feedback; if the staged library exceeds size or length limits, the manager reorganizes it before acceptance. When a proposal is accepted, its changes are committed and the insights it absorbs are marked as reviewed, excluding them from subsequent updates so that each update draws only on new evidence. Harness versions. Frozen skill-only replay and evaluation disable cross-task insight retrieval and ASE updates. Accumulated experience therefore reaches the agent only through the committed skill library, so each committed update defines a new harness version. We denote by ℋrH_r the harness equipped with the r-th skill library rS_r. Among the resulting versions, the one to internalize with D-OPCD is selected on the held-out task set holdQ_hold. 3.4 Harness–Generator Co-Evolution Harness evolution improves the agent without training, but the harness has limited capacity: each added skill consumes context and makes it harder for the agent to choose the right one at inference time. D-OPCD relieves this limit by internalizing part of the harness knowledge into the generator weights. After the generator update, we clear the learned skills to examine whether the harness can acquire new guidance around the updated generator. Data protocol. For each benchmark, we split the tasks into three disjoint sets, =train⊔hold⊔evalQ=Q_train _hold _eval. trainQ_train is used for harness evolution and D-OPCD training; holdQ_hold only for selecting harness versions and training checkpoints; and evalQ_eval only for reporting final results. Harness evolution and version selection. With the current generator πθ _θ fixed, the agent processes trainQ_train as a stream and evolves its skill library with ASE (Section 3.3), producing harness versions ℋr\H_r\. We evaluate each version on holdQ_hold and select the best-performing version ℋr⋆H_r using the external evaluator’s aggregate score J. r⋆=arg⁡maxr⁡J⁡(ℋr,πθ,hold).r = _r\ J (H_r;\ _θ,\ Q_hold ). (8) Replay and internalization. We freeze the skill library of ℋr⋆H_r and replay it on trainQ_train with πθ _θ. The resulting records form ctx(r⋆)D_ctx^(r ) in Equation 3, and D-OPCD uses this dataset to update the generator to πθ′ _θ . We replay rather than reuse the streaming records so that all contexts come from a single harness version operating on the generator from which D-OPCD is initialized. Generator adoption and harness reset. After D-OPCD, we adopt πθ′ _θ as the new generator. To measure how much of the harness gain has been transferred into the weights, we define the signed internalization gap δ(r⋆)=J⁡(ℋ∅,πθ′)−J⁡(ℋr⋆,πθ).δ^(r )=J (H_ ; _θ )-J (H_r ; _θ ). (9) Both scores are measured on evalQ_eval. A value near zero means that the updated generator without learned skills matches the original skill-equipped agent; a positive value favors the updated generator, while a negative value indicates a remaining deficit. We clear the skill library and its associated episode and insight records, retain the skill-free harness ℋ∅H_ , and evolve new harness versions against πθ′ _θ . We denote by ℋr†H_r a version selected on holdQ_hold under πθ′ _θ after this renewed harness evolution. 4 Experiments 4.1 Experimental Setup Tasks and benchmarks. We evaluate on GenEval (Ghosh et al., 2023), GenEval2 (Kamath et al., 2025), WISE (Niu et al., 2025), and R2I-Bench (Chen et al., 2025), which cover compositional, knowledge-intensive, and reasoning-intensive generation; T2I-CompBench++ (Huang et al., 2025) is reserved for out-of-distribution (OOD) evaluation. Each benchmark forms a sequential task stream: the agent receives a query and obtains evaluator feedback after submitting its final image. We use trainQ_train, holdQ_hold, and evalQ_eval as defined in Section 3.4, and report evaluation results on evalQ_eval. The official evaluators provide GenEval Overall, GenEval2 Soft-TIFA AM, WIScore, and R2I-Score, respectively. Training-method comparisons use direct generation from qiq_i; the main results also evaluate the base and D-OPCD-updated generators with skill-free and evolved harnesses. Appendix E.1 gives split sizes, and Appendix E details the evaluation protocols. Agent and harness evolution. We use Z-Image-Turbo as the initial T2I generator. The agent starts from a GEMS-based scaffold (He et al., 2026) with an empty skill library (the skill-free harness) and is equipped with the Auto Skill Evolver (ASE) described in Section 3.3. With the generator fixed, it processes trainQ_train as a task stream; ASE extracts insights from completed episodes and consolidates mature insights into skills, producing successive harness versions. Appendix C.1 describes the scaffold design, and Appendix C.4 reports the benchmark-specific ASE settings. Compared training methods. For a matched comparison, all methods use the same prompt-changed task records replayed from the selected skill-equipped harness ℋr⋆H_r on trainQ_train. They start from the same Z-Image-Turbo weights and train LoRA adapters with the same optimization budget. D-OPCD uses each (qi,pi)(q_i,p_i) pair: an EMA teacher conditioned on [qi;pi][q_i;p_i] supervises a student conditioned only on qiq_i along the student’s own sampling trajectory, with pip_i hidden at inference. Vanilla SFT fits the agent-submitted image IiI_i from qiq_i with a standard flow-matching objective. Diffusion-DPO (Wallace et al., 2024) treats IiI_i as chosen and a same-seed image generated directly from qiq_i as rejected, training both under qiq_i. D-OPSD (Jiang et al., 2026a) uses an image-conditioned teacher based on (qi,Ii)(q_i,I_i) to supervise a qiq_i-conditioned student along its own trajectory. Appendix D.1 details the record selection; Appendix D gives training objectives and settings. 4.2 Main Results Table 1: Base and D-OPCD generator scores under direct generation, a skill-free harness, and evolved skills. Parenthesized Δ values compare base ++ rows with base direct generation and D-OPCD rows with the corresponding rows in the base-generator block above, in the same order. Method GenEval ↑ GenEval2 ↑ WISE ↑ R2I-Bench ↑ Avg. ↑ Base generator, direct 78.22 78.14 42.86 42.87 60.52 + skill-free harness (ℋ∅H_ ) 83.51 (+5.29) 87.50 (+9.36) 78.00 (+35.14) 68.12 (+25.25) 79.28 (+18.76) + evolved skills (ℋr⋆H_r ) 87.63 (+9.41) 88.83 (+10.69) 81.43 (+38.57) 69.41 (+26.54) 81.83 (+21.31) D-OPCD generator, direct 81.28 (+3.06) 82.24 (+4.10) 50.00 (+7.14) 46.83 (+3.96) 65.09 (+4.57) + skill-free harness (ℋ∅H_ ) 86.08 (+2.57) 88.41 (+0.91) 81.14 (+3.14) 73.68 (+5.56) 82.33 (+3.05) + newly evolved skills (ℋr†H_r ) 91.75 (+4.12) 89.74 (+0.91) 85.71 (+4.28) 69.45 (+0.04) 84.16 (+2.33) Harness gains at the generator interface. To assess the gains provided by the agent harness, we evaluate the base generator πθ _θ on evalQ_eval alone, with the skill-free harness ℋ∅H_ , and with the selected skill-equipped harness ℋr⋆H_r evolved by ASE. Both harness configurations improve performance over direct generation on all four benchmarks (Table 1). The skill-free harness raises the four-benchmark average from 60.52 to 79.28 (+18.76 points), including gains from 42.86 to 78.00 on WISE and from 42.87 to 68.12 on R2I-Bench. ASE-evolved skills raise the average further to 81.83 (+2.55 points over ℋ∅H_ ), with additional gains on every benchmark, ranging from +1.29 points on R2I-Bench to +4.12 on GenEval. The prompts selected by ℋr⋆H_r then provide the privileged contexts for D-OPCD. Internalizing harness gains. To test whether the harness’s gains can be internalized, we train the generator on (qi,pi)(q_i,p_i) pairs from ℋr⋆H_r with D-OPCD and evaluate πθ′ _θ directly from qiq_i, without a harness. It outperforms the base generator on all four benchmarks, raising the average score from 60.52 to 65.09 (Table 1). Some of the harness benefit therefore persists in the model weights. Harness benefits after internalization. To test whether internalization reduces reliance on the original evolved skills, we attach the skill-free harness ℋ∅H_ to πθ′ _θ and compare it with the base generator under ℋ∅H_ and ℋr⋆H_r . The updated skill-free agent outperforms its base-generator counterpart on all four benchmarks (Table 1). It scores 82.33 on average and slightly exceeds the original skill-equipped agent, yielding a signed internalization gap of δ(r⋆)=+0.50δ^(r )=+0.50. On R2I-Bench, it is also ahead, with δ(r⋆)=+4.27δ^(r )=+4.27. Thus, much of the earlier skill benefit can be retained after removing those skills, although the gap varies across benchmarks. Continued co-evolution after model update. After adopting πθ′ _θ , we reset the learned skills and rerun ASE to obtain ℋr†H_r . The renewed skills improve the updated agent on three benchmarks and raise its average above both the updated skill-free agent and the original skill-equipped agent (Table 1). R2I-Bench instead declines after skill evolution. The D-OPCD generator with the skill-free harness already scores highly, which may leave less room for additional gains. These results illustrate how D-OPCD unlocks a harness–model co-evolution mechanism, with renewed skills building on the updated generator. Matched image examples are provided in Appendix A. 4.3 Ablation Studies and Further Analysis 4.3.1 Comparison with Alternative Training Methods To compare methods for internalizing agent experience, we train D-OPCD and three baselines on matched records collected by the evolved harness and evaluate direct generation from the original queries (Table 2). All four improve the base generator’s four-benchmark average, with gains ranging from 1.66 to 4.57 points. D-OPCD achieves the highest mean of 65.09, 1.71 points above Vanilla SFT, and leads on all four benchmarks. On WISE, D-OPCD improves over the base generator by 7.14 points. Overall, these results suggest that D-OPCD has an advantage over the tested baselines in internalizing agent experience into the generator’s model weights. Table 2: Direct-generation comparison on evalQ_eval. Updated models train on matched records from ℋr⋆H_r : D-OPCD uses (qi,pi)(q_i,p_i), while baselines use the agent image IiI_i. Parenthesized Δ values are differences from the base generator; Avg. is the unweighted mean across four benchmarks. Method GenEval ↑ GenEval2 ↑ WISE ↑ R2I-Bench ↑ Avg. ↑ Base generator 78.22 78.14 42.86 42.87 60.52 Vanilla SFT 80.09 (+1.87) 80.94 (+2.80) 46.00 (+3.14) 46.51 (+3.64) 63.38 (+2.86) Diffusion-DPO (Wallace et al., 2024) 77.56 (-0.66) 79.04 (+0.90) 45.43 (+2.57) 46.70 (+3.83) 62.18 (+1.66) D-OPSD (Jiang et al., 2026a) 80.72 (+2.50) 76.48 (-1.66) 48.86 (+6.00) 44.96 (+2.09) 62.75 (+2.23) D-OPCD (ours) 81.28 (+3.06) 82.24 (+4.10) 50.00 (+7.14) 46.83 (+3.96) 65.09 (+4.57) 4.3.2 Ablation of Teacher Conditioning With the student conditioned on qiq_i, we compare three teacher inputs: matched [qi;pi][q_i;p_i], pip_i alone, and shuffled [qi;pσ⁡(i)][q_i;p_σ(i)] using another task’s prompt. The matched input averages 65.09, ahead of pip_i alone (64.63) and shuffled prompts (61.05). It leads on three benchmarks, while pip_i alone reaches the highest R2I-Bench score of 47.32. Shuffling reduces scores on all four benchmarks, underscoring the value of task-matched teacher prompts (Table 3). Table 3: D-OPCD performance with different teacher inputs on evalQ_eval. The teacher receives the original query with its matched harness prompt ([qi;pi][q_i;p_i]), the harness prompt alone (pip_i only), or the query with a prompt from another task (Shuffled). The student always receives qiq_i. Teacher input GenEval ↑ GenEval2 ↑ WISE ↑ R2I-Bench ↑ Avg. ↑ D-OPCD: [qi;pi][q_i;p_i] 81.28 82.24 50.00 46.83 65.09 pip_i only 80.93 81.13 49.14 47.32 64.63 Shuffled: [qi;pσ⁡(i)][q_i;p_σ(i)] 74.50 75.38 48.57 45.76 61.05 4.3.3 Context Properties and Model Internalization Table 1 shows that training on contexts replayed from the selected skill-equipped harness ℋr⋆H_r improves direct generation from the original query on all four benchmarks. To explore how context properties relate to model internalization, we examine two observable properties of the collected (qi,pi)(q_i,p_i) pairs: the submitted image’s advantage over direct-query generation and variation in prompt wording. We compare these data-level properties with direct-generation scores after training. Context design. We vary the default prompt-changed replay set in two ways. 1) Improvement filter. We retain a pair only when the agent’s internal verifier gives the submitted image a strictly higher pass count than the direct-query image. 2) Diversity adapter. We change the agent’s prompt-sampling policy to reduce repeated prompt patterns, then collect its prompt-changed pairs. Appendix E.3 gives the construction details for both variants. Results and analysis. The improvement-filtered pairs show larger observed image-score gaps on their retained tasks, while the diversity adapter increases character-level variation in the collected prompts (Appendix E.3). However, downstream gains are mixed: both variants improve individual benchmarks, but the filter leaves the four-benchmark mean nearly unchanged and the adapter lowers it (Table 4). Filtering also changes the training set’s size and composition, so its effect cannot be attributed solely to task improvement. A prompt that is more useful during generation is therefore not necessarily easier for the query-only student to internalize when used as context. Identifying which harness-produced contexts best support model internalization remains an open question. Table 4: D-OPCD direct-generation test scores after training on three context sets. Pair counts follow GenEval/GenEval2/WISE/R2I-Bench order; Avg. is the four-benchmark mean before rounding. Setting Pairs GenEval ↑ GenEval2 ↑ WISE ↑ R2I-Bench ↑ Avg. ↑ Default (p≠qp≠ q) 276/400/490/392 81.28 82.24 50.00 46.83 65.09 Improvement filter 43/264/319/314 82.98 81.76 47.71 48.13 65.14 Diversity adapter 276/400/491/381 82.74 81.75 48.29 44.92 64.42 4.3.4 Out-of-Distribution Generalization We evaluate the base generator and four source-specific D-OPCD checkpoints through direct generation without a harness on seven T2I-CompBench++ categories reserved for out-of-distribution testing (Huang et al., 2025). Color scores rise from 77.45 for the base generator to 81.61 and 81.44 for the GenEval- and GenEval2-trained checkpoints, respectively. Their 2D-Spatial scores likewise reach 36.19 and 37.33, up from 32.32 for the base (Table 5). The WISE- and R2I-Bench-trained checkpoints remain close to the base overall. These results show good out-of-distribution performance, especially with GenEval and GenEval2 as training sources. Table 5: Out-of-distribution category scores on T2I-CompBench++ for the base generator and four D-OPCD checkpoints. The suffix in each D-OPCD method name identifies its training benchmark. Method Color Shape Texture 2D-Spatial 3D-Spatial Numeracy Non-Spatial Base generator 77.45 57.54 72.79 32.32 41.27 67.98 31.24 D-OPCD-GenEval 81.61 57.47 73.00 36.19 41.53 70.26 31.52 D-OPCD-GenEval2 81.44 56.56 72.36 37.33 42.02 71.05 31.31 D-OPCD-WISE 77.17 57.21 72.67 32.34 39.74 67.09 31.34 D-OPCD-R2I-Bench 77.06 57.84 73.09 32.00 41.07 68.42 31.33 5 Conclusion We introduce D-OPCD to internalize prompt-mediated gains from an evolving T2I harness into the generator’s weights, allowing the model to retain part of the harness’s benefit from the original query alone. With ASE turning task experience into reusable skills, D-OPCD improves direct generation across all four benchmarks. The updated generator with a skill-free harness slightly exceeds the original skill-equipped agent on average, and resetting the old skills allows ASE to evolve again and further improve performance. We hope this work guides the development of T2I systems in which accumulated experience becomes lasting generator capability, creating room for the harness and model to keep improving each other through continual co-evolution. References Black et al. (2024) K. Black, M. Janner, Y. Du, I. Kostrikov, and S. Levine Training diffusion models with reinforcement learning. In International Conference on Learning Representations, B. Kim, Y. Yue, S. Chaudhuri, K. Fragkiadaki, M. Khan, and Y. Sun (Eds.), Vol. 2024, p. 4965–4987. External Links: Link Cited by: §2.2. Chen et al. (2025) K. Chen, Z. Lin, Z. Xu, Y. Shen, Y. Yao, J. Rimchala, J. Zhang, and L. Huang R2I-bench: benchmarking reasoning-driven text-to-image generation. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, C. Christodoulopoulos, T. Chakraborty, C. Rose, and V. Peng (Eds.), Suzhou, China, p. 12595–12630. External Links: Link, Document, ISBN 979-8-89176-332-6 Cited by: §E.4, §4.1. Chen et al. (2026a) S. Chen, Z. Xing, T. Ye, X. Geng, Y. Lin, J. Lai, X. He, F. Zhai, J. Gao, and L. Zhu GenEvolve: self-evolving image generation agents via tool-orchestrated visual experience distillation. arXiv. External Links: Document, Link Cited by: §1, §2.1. Chen et al. (2026b) W. Chen, K. Yu, B. Tian, J. Song, S. Liang, H. Jia, K. Cheng, H. Li, K. Yuan, L. Wang, J. Wu, S. Lai, and Y. Yue MemoGen: can past experience improve future text-to-image generation?. arXiv. External Links: Document, Link Cited by: §1, §2.1. Clark et al. (2024) K. Clark, P. Vicol, K. Swersky, and D. Fleet Directly fine-tuning diffusion models on differentiable rewards. In International Conference on Learning Representations, B. Kim, Y. Yue, S. Chaudhuri, K. Fragkiadaki, M. Khan, and Y. Sun (Eds.), Vol. 2024, p. 4793–4822. External Links: Link Cited by: §2.2. Fang et al. (2026) Z. Fang, W. Huang, Y. Zeng, Y. Zhao, S. Chen, K. Feng, Y. Lin, L. Chen, Z. Chen, S. Cao, and F. Zhao Flow-opd: on-policy distillation for flow matching models. arXiv. External Links: Document, Link Cited by: §2.2. Friedman and Dieng (2023) D. Friedman and A. B. Dieng The vendi score: a diversity evaluation metric for machine learning. Transactions on Machine Learning Research. External Links: ISSN 2835-8856 Cited by: §E.3. Ghosh et al. (2023) D. Ghosh, H. Hajishirzi, and L. Schmidt GenEval: an object-focused framework for evaluating text-to-image alignment. In Advances in Neural Information Processing Systems 36, NeurIPS 2023, p. 52132–52152. External Links: Link, Document Cited by: §E.4, §4.1. Hao et al. (2023) Y. Hao, Z. Chi, L. Dong, and F. Wei Optimizing prompts for text-to-image generation. In Advances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (Eds.), Vol. 36, p. 66923–66939. External Links: Document, Link Cited by: §2.1. He et al. (2026) Z. He, S. Huang, X. Qu, Y. Li, T. Zhu, Y. Cheng, and Y. Yang GEMS: agent-native multimodal generation with memory and skills. arXiv. External Links: Document, Link Cited by: §C.1, §1, §2.1, §4.1. Huang et al. (2025) K. Huang, C. Duan, K. Sun, E. Xie, Z. Li, and X. Liu T2I-compbench++: an enhanced and comprehensive benchmark for compositional text-to-image generation. IEEE Transactions on Pattern Analysis and Machine Intelligence 47 (5), p. 3563–3579. External Links: ISSN 1939-3539, Link, Document Cited by: §E.4, §4.1, §4.3.4. Jiang et al. (2026a) D. Jiang, X. Jin, D. Liu, Z. Wang, M. Zheng, R. Du, X. Yang, Q. Wu, Z. Li, P. Gao, H. Yang, and S. Hoi D-opsd: on-policy self-distillation for continuously tuning step-distilled diffusion models. arXiv. External Links: Document, Link Cited by: §1, §2.2, §3.2.1, §3.2.2, §4.1, Table 2. Jiang et al. (2026b) K. Jiang, Y. Wang, J. Zhou, P. Li, Z. Liu, C. Xie, Z. Chen, Y. Zheng, and W. Zhang GenAgent: scaling text-to-image generation via agentic multimodal reasoning. arXiv. External Links: Document, Link Cited by: §2.1. Kamath et al. (2025) A. Kamath, K. Chang, R. Krishna, L. Zettlemoyer, Y. Hu, and M. Ghazvininejad GenEval 2: addressing benchmark drift in text-to-image evaluation. arXiv. External Links: Document, Link Cited by: §E.4, §4.1. Karten et al. (2026) S. Karten, J. Zhang, T. Upaa, R. Feng, W. Li, C. Shi, C. Jin, and K. Vodrahalli Continual harness: online adaptation for self-improving foundation agents. arXiv. External Links: Document, Link Cited by: §2.3. Kovalev et al. (2025) V. Kovalev, A. Kuvshinov, A. Buzovkin, D. Pokidov, and D. Timonin CRAFT: continuous reasoning and agentic feedback tuning for multimodal text-to-image generation. arXiv. External Links: Document, Link Cited by: §2.1. Li et al. (2016) J. Li, M. Galley, C. Brockett, J. Gao, and B. Dolan A diversity-promoting objective function for neural conversation models. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, p. 110–119. External Links: Link, Document Cited by: §E.3. Liu et al. (2025) J. Liu, G. Liu, J. Liang, Y. Li, J. Liu, X. Wang, P. Wan, D. Zhang, and W. Ouyang Flow-grpo: training flow matching models via online rl. arXiv. External Links: Document, Link Cited by: §2.2. Niu et al. (2025) Y. Niu, M. Ning, M. Zheng, W. Jin, B. Lin, P. Jin, J. Liao, C. Feng, F. Meng, K. Ning, B. Zhu, and L. Yuan WISE: a world knowledge-informed semantic evaluation for text-to-image generation. arXiv. External Links: Document, Link Cited by: §E.4, §4.1. Ruiz et al. (2023) N. Ruiz, Y. Li, V. Jampani, Y. Pritch, M. Rubinstein, and K. Aberman DreamBooth: fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 22500–22510. Cited by: §2.2. Smith et al. (2023) J. S. Smith, Y. Hsu, L. Zhang, T. Hua, Z. Kira, Y. Shen, and H. Jin Continual diffusion: continual customization of text-to-image diffusion with c-lora. arXiv. External Links: Document, Link Cited by: §2.2. Wallace et al. (2024) B. Wallace, M. Dang, R. Rafailov, L. Zhou, A. Lou, S. Purushwalkam, S. Ermon, C. Xiong, S. Joty, and N. Naik Diffusion model alignment using direct preference optimization. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 8228–8238. External Links: Link, Document Cited by: §2.2, §4.1, Table 2. Wang et al. (2025) K. Wang, R. Chen, T. Zheng, and H. Huang ImAgent: a unified multimodal agent framework for test-time scalable image generation. arXiv. External Links: Document, Link Cited by: §1, §2.1. Wu et al. (2026) J. Wu, S. Yang, Z. Lu, F. Zhang, Y. Shen, L. Feng, H. Luo, Z. Lian, S. Zhang, Z. Wen, and J. Tao SEED: self-evolving on-policy distillation for agentic reinforcement learning. arXiv. External Links: Document, Link Cited by: §2.3. Xia et al. (2026a) P. Xia, J. Chen, H. Wang, J. Liu, K. Zeng, Y. Wang, S. Han, Y. Zhou, X. Zhao, H. Chen, Z. Zheng, C. Xie, and H. Yao SkillRL: evolving agents via recursive skill-augmented reinforcement learning. arXiv. External Links: Document, Link Cited by: §2.3. Xia et al. (2026b) P. Xia, J. Chen, X. Yang, H. Tu, J. Liu, K. Xiong, S. Han, S. Qiu, H. Ji, Y. Zhou, Z. Zheng, C. Xie, and H. Yao MetaClaw: just talk – an agent that meta-learns and evolves in the wild. arXiv. External Links: Document, Link Cited by: §2.3. Xu et al. (2023) J. Xu, X. Liu, Y. Wu, Y. Tong, Q. Li, M. Ding, J. Tang, and Y. Dong ImageReward: learning and evaluating human preferences for text-to-image generation. In Advances in Neural Information Processing Systems 36, NeurIPS 2023, p. 15903–15935. External Links: Link, Document Cited by: §2.2. Ye et al. (2026) T. Ye, L. Dong, X. Wu, S. Huang, and F. Wei On-policy context distillation for language models. arXiv. External Links: Document, Link Cited by: §1, §3.2.1. Zhang et al. (2026a) H. Zhang, Q. Long, J. Bao, T. Feng, W. Zhang, H. Yue, and W. Wang MemSkill: learning and evolving memory skills for self-evolving agents. arXiv. External Links: Document, Link Cited by: §2.3. Zhang et al. (2026b) Z. Zhang, J. Li, J. Zhang, K. Gao, K. Yan, L. Jiang, N. Tang, S. Yin, T. Wu, X. Chen, X. Xu, Y. Shu, Y. Zhang, Y. Xu, Y. Chen, Z. Wang, Z. Liu, Z. Zhou, H. Zhang, D. Zhao, and C. Wu Qwen-image-agent: bridging the context gap in real-world image generation. arXiv. External Links: Document, Link Cited by: §2.1. Zhang et al. (2026c) Z. Zhang, Y. Lin, N. L. Kuang, L. Wu, X. Li, S. Liu, and F. Ma Co-evolving skill generation and policy optimization. arXiv. External Links: Document, Link Cited by: §2.3. Zhao et al. (2024) A. Zhao, D. Huang, Q. Xu, M. Lin, Y. Liu, and G. Huang ExpeL: llm agents are experiential learners. Proceedings of the AAAI Conference on Artificial Intelligence 38 (17), p. 19632–19642. External Links: ISSN 2159-5399, Link, Document Cited by: §2.3. Appendix A Qualitative Examples for the Main Results Figure 3 shows selected test queries for which D-OPCD improves direct generation. Table 1 reports the aggregate results over the full evaluation sets. Figure 3: Matched outputs for four test queries. The left triplet uses the base generator πθ _θ and the right triplet uses the D-OPCD-updated generator πθ′ _θ . Within each triplet, the columns show direct generation, the skill-free harness ℋ∅H_ , and evolved skills (ℋr⋆H_r or ℋr†H_r ). The text below each row is the original query; harness runs may use selected generation prompts. Appendix B Discussion on Limitations and Future Works Limits of internalizing skill-evolution experience. Internalizing the experience from skill evolution yields only modest gains for the standalone generator: D-OPCD raises the four-benchmark direct-generation average from 60.52 to 65.09, still far below the 81.83 achieved by the original skill-equipped harness (Table 1). One possible explanation is that prompts produced by evolved skills, while effective when used by the harness, may be poorly aligned with the diffusion model’s latent space for learning the same behavior from the original query qiq_i alone. Further studies could explore which harness designs produce experience that supports both immediate task success and subsequent model learning, and which algorithms can internalize this broader harness experience. Stronger mechanisms for jointly adapting the harness and generator may be needed to realize sustained co-evolution. Evidence for repeated co-evolution. Our experiments evaluate one base generator and one D-OPCD update on four benchmarks, followed by one subsequent round of skill evolution. They show that a newly evolved harness can build on the updated generator, but do not establish what happens over many update-and-reset cycles or across different generators and task distributions. Longer-term studies could measure retention of earlier gains, transfer to new tasks, and the tradeoff between offline training cost and recurring harness inference cost. Repeated model updates may also require methods that prevent forgetting while adapting the harness to each new generator. Beyond text-to-image agents. The agent studied here improves generation through prompt construction, verification, and refinement, without external knowledge search or reference-image tools. Other image-generation agents may search the web, retrieve reference images, and use additional tools. In addition to text-to-image generation, they may handle image-to-image generation and other image-generation tasks. Their experience includes tool choices, retrieved information, and visual evidence that a final text prompt cannot fully represent. Extending internalization to these settings requires learning which reusable capabilities can be transferred into model weights and which information should remain available through tools at inference. A further open problem is how to use tool-augmented experience gathered at test time to guide model updates and then evolve the harness again around the updated generator. Appendix C Implementation Details of the Agent Harness and Skill Evolution C.1 GEMS-Based Agent Harness Design Planning and requirements. We adapt the GEMS agent architecture (He et al., 2026) to a text-to-image harness with five roles: Planner, Decomposer, Generator, Verifier, and Refiner. Given an original query qiq_i, the Planner selects applicable Skills using their names and descriptions, then loads their instructions to construct the first generation prompt pi,1p_i,1. If it selects no Skill, pi,1=qip_i,1=q_i. The Decomposer then turns qiq_i into a fixed set of yes/no visual requirements i=(ci,1,…,ci,mi)C_i=(c_i,1,…,c_i,m_i). Deriving these checks from the original query keeps the target of verification fixed as prompts are revised. Generation and refinement. The Generator produces an image Ii,jI_i,j from the current prompt pi,jp_i,j. The multimodal Verifier checks the image against every requirement in iC_i. If any check fails, the Refiner uses the prompt, image, verification feedback, and task-local history of earlier attempts to produce pi,j+1p_i,j+1. That history retains the prompts, images, check results, and short experience summaries used during refinement. The agent stops when all checks pass or the attempt budget is exhausted. During evolution, relevant Insights may also be retrieved to guide refinement after a failure; frozen skill-only replay and evaluation disable this retrieval. External benchmark scores do not guide refinement within a task. Returned output. Let ni,jn_i,j be the number of requirements satisfied by attempt j, and let JiJ_i be the number of attempts executed. The agent returns the prompt and image from the earliest attempt with the highest count: ji∗=min⁡j:ni,j=max1≤ℓ≤Ji⁡ni,ℓ,(pi,Ii)=(pi,ji∗,Ii,ji∗).j_i^*= \j:n_i,j= _1≤ ≤ J_in_i, \, (p_i,I_i)=(p_i,j_i^*,I_i,j_i^*). (10) Thus, a later unsuccessful revision can leave an earlier image as the returned output. The resulting (pi,Ii)(p_i,I_i) is the harness output used for context-data collection. The agent prompt templates are listed in Appendix C.5. C.2 Episode Construction and Insight Extraction Episode records. Each completed task yields one Episode. An Episode records (i) the task metadata, including the original request qiq_i and the deployed generator version; (i) the ordered generation attempts and verification results; (i) the Skills selected at the start of the task and any Insights retrieved for refinement; and (iv) the returned attempt, its passed and failed checks, the first-attempt check score, and any feedback from the external evaluator. Multiple attempts within the same Episode still count as one evidence unit. An optional LLM-written summary may be attached, but we disable it so that consolidation uses the factual attempt record and outcome. Frozen working set. At the batch barrier, the learner builds one retrieval query per Episode from its request and failure evidence, retrieves the top-44 active Insights whose cosine similarity to the query exceeds 0.350.35 using BGE-M3 embeddings, and takes the union over Episodes as the working set kN_k. This management-only retrieval includes immature single-Episode hypotheses, so that new evidence can strengthen, correct, or refute them. The working set is frozen before the manager runs: the Insight manager ℳIM_I may reference only these entries and only Episodes of the current batch. Action space. ℳIM_I returns an ordered list of operations. Table 6 summarizes the six actions, and Figure 10 shows the manager’s decision prompt. The action space is deliberately small: evidence moves are separated from text moves, so support and contradiction accumulate as Episode references while wording changes remain rare, explicit, and evidence-backed. Table 6: Action space of the Insight manager ℳIM_I. All actions require at least one Episode reference from the current batch; targets must be active Insights in the frozen working set. Action Parameters Semantics Key validation rules Add text, episode_refs Create an Insight capturing a distinct, transferable refinement lesson. Exact or semantic (≥0.92≥ 0.92 cosine) duplicates are not created; the action is resolved as Support for the matching Insight instead. Support insight_ref, episode_refs Attach supporting Episodes without changing the text. Cannot carry text; the same Episode supports an Insight at most once per batch. Revise insight_ref, text, episode_refs Refine the scope, boundary, or wording of an Insight and attach the motivating support. Requires non-empty text; an Insight may be revised at most once per batch. Contradict insight_ref, episode_refs, reason Record Episode evidence that weakens or limits an Insight. Requires a reason; one Episode cannot both support and contradict the same Insight. Merge target_ref, merged_refs, text, episode_refs Fold overlapping Insights into the target with a complete reformulation; sources are archived. Target and sources must be distinct active entries; the merged evidence is re-attributed to the target. Archive insight_ref, episode_refs, reason, [replacement_ref] Retire obsolete or redundant knowledge, optionally pointing to a replacement Insight. Requires a reason; the replacement must be active and different from the archived entry. Validation and commit. Operations are checked in order against the frozen working set: references must resolve, text constraints must hold per action, and one Episode may not take both a supporting and a contradicting direction toward the same Insight. Only operations satisfying these rules update the Insight pool. Evidence and maturity. Support and contradiction counts use distinct Episodes, not the number of attempts within an Episode. The support threshold τ and the minimum number of supporting batches required before an Insight is passed to the Skill manager vary by benchmark (Table 9). The Insight store retains contradictory Episode references so they can affect maturity and motivate a revision or archival. C.3 Skill Updates and Library Versioning Update trigger. After each batch k, Skill evolution is considered when at least B mature, previously unreviewed Insights have accumulated and the minimum interval since the preceding completed Skill review has elapsed. Both settings vary by benchmark (Table 9). The mature-insight trigger controls Skill maintenance independently of generator updates. Action space. Once triggered, the Skill manager ℳSM_S examines the pending mature Insights together with summaries of the current library and the recent change history, and returns one ordered plan from the nine-action space in Table 7. Figure 11 shows its decision prompt. Text-level edits (Replace, Insert, Delete) operate on exact, uniquely matching passages copied from the target Skill, which makes every proposed diff verifiable before it is applied. Structural actions (Split, Merge, Retire) keep the library routeable under a hard capacity bound. Every edit cites the Insights it absorbs, linking each Skill revision back to its evidence. Table 7: Action space of the Skill manager ℳSM_S. source_insight_refs identifies the mature Insights absorbed by an edit. An empty plan is a deliberate no-op. Action Parameters Semantics Key validation rules Create suggested_id, content, source_insight_refs Add a distinct high-level Skill as complete natural-language content. Requires at least one source Insight; subject to the capacity bound C. ReplaceText skill_id, old_text, new_text, source_insight_refs Replace an exact passage in one Skill. old_text must match a unique location in the target Skill. InsertText skill_id, anchor, text, position, source_insight_refs Insert text immediately before or after an anchor. The anchor must match uniquely; position∈before,after position∈\ before, after\. DeleteText skill_id, text, source_insight_refs Remove an exact passage from one Skill. The passage must match uniquely. Rename skill_id, new_title, source_insight_refs Change a Skill’s title while preserving its stable identifier and body. Requires a non-empty title. Summary skill_id, content, source_insight_refs Replace an overgrown Skill with shorter complete content, preserving its identity. Result must satisfy the length bound L. Split skill_id, new_skills, source_insight_refs Replace one over-broad Skill with two complete, independently routeable Skills. Exactly two replacement Skills; subject to C. Merge target_skill_id, merged_skill_ids, content, source_insight_refs Merge overlapping Skills into the target with complete resulting content. Sources are retired; result must satisfy C and L. Retire skill_id, replacement_skill_id Retire a redundant Skill in favor of an active replacement. The replacement must be an active Skill. Validation and commit. Proposed plans are checked against two hard constraints: at most C=5C=5 active Skills, and at most L=4500L=4500 characters per Skill. Only a valid final library is committed. Each committed edit records the Insights it absorbs. Reviewed evidence and harness versions. The review state is recorded per Insight at its latest evidence revision. An accepted Skill update marks the Insights it absorbs as reviewed, preventing the same unchanged evidence from immediately triggering another update. A deliberate NoOp also completes a review but changes neither the active library nor its version. If a reviewed Insight later gains new evidence, its new revision can be considered in a subsequent update. Each committed library change defines a new harness version. Deployment-time use. At inference the agent routes each new request to the relevant Skills once, before the first generation. Generation calls under the evolving library are recorded with their harness and generator version annotations and become the prompt–image pairs available to D-OPCD. C.4 Experimental Configuration and Frozen Replay Experimental configuration. GPT-5.6 Terra with medium reasoning effort serves as the agent’s base reasoning model, the Insight and Skill managers, and the in-task verifier. Table 8 reports settings shared across benchmarks; Table 9 reports benchmark-specific ASE settings. Evaluation leaves the agent state unchanged, so it cannot affect subsequent predictions or the adaptation stream. Table 8: Harness-evolution settings shared across benchmarks. Component Parameter Value Agent loop max verification questions 1010 Agent loop max generation attempts 55 Episode store LLM summary disabled Insight extraction task batch size M 88 Insight retrieval text embedder BGE-M3 Insight retrieval top-k / semantic gate 44 / 0.350.35 cosine Insight consolidation duplicate threshold 0.920.92 cosine Skill library max active Skills C 55 Skill library max Skill length L 45004500 characters Table 9: Benchmark-specific ASE settings during evolution on trainQ_train. The review interval is measured in task batches. Parameter GenEval GenEval2 WISE R2I-Bench Insight support τ 22 44 44 44 Min. support batches 22 33 33 33 Skill trigger B 33 88 88 88 Min. review interval 44 66 66 66 Generation seeds. For task i, generation attempt j uses seed si+j−1s_i+j-1, where sis_i is the task’s base seed. Selecting and freezing a skill library. Library selection is performed on holdQ_hold, independently of evaluation on evalQ_eval. Each candidate is evaluated without persistent updates; the selected library is subsequently frozen for replay on the entire trainQ_train and for evaluation on evalQ_eval. The replayed tasks do not update the Skill library or the persistent Insight store. Table 10 identifies which Skill-update snapshot was selected in each evolution round and how many tasks had contributed experience to that snapshot. The skill-free harness ℋ∅H_ uses the empty library and requires no version selection. Table 10: Selected Skill-update snapshots in both evolution rounds. The selected update is r⋆r in Round 1 and r†r in Round 2; nrn_r counts completed tasks contributing experience to that snapshot. Selection scores use GenEval2 Soft-TIFA AM and the other benchmarks’ native metrics. Evolution round Benchmark Selected Skill update r Experience from nrn_r tasks Selection score ↑ Round 1 (πθ _θ) GenEval 2 192 92.77 Round 1 (πθ _θ) GenEval2 3 352 90.85 Round 1 (πθ _θ) WISE 4 280 82.67 Round 1 (πθ _θ) R2I-Bench 4 216 69.60 Round 2 (πθ′ _θ ) GenEval 2 240 92.77 Round 2 (πθ′ _θ ) GenEval2 1 128 90.40 Round 2 (πθ′ _θ ) WISE 3 272 82.00 Round 2 (πθ′ _θ ) R2I-Bench 4 328 68.44 Prompt-pair collection. For each query qi∈trainq_i _train, frozen replay returns (pi,Ii)(p_i,I_i) according to Equation 10. We retain the submitted record (qi,pi,Ii)(q_i,p_i,I_i); all compared training methods drop tasks whose prompt equals the query after normalization, following the context-data protocol in Appendix D.1. C.5 Prompt Templates C.5.1 Agent Execution Prompts Figures 4–9 show the prompts used by the agent during each generation task in execution order. Braced capitalized fields denote runtime substitutions, and line wrapping is for presentation only. The optional cross-task Insight field in the refinement template is used during evolution and left empty in frozen skill-only replay and evaluation. ⬇ You are an image-generation Skill Router. Every available Skill is an initial-prompt rewriter that may run exactly once before the first image generation. Choose the subset of candidate Skills whose transformations are applicable and useful for the original request. You may combine multiple complementary Skills or choose none. ### Candidate Initial-Prompt Skills SKILL_MANIFEST ### Original User Request ORIGINAL_QUERY Return ONLY a JSON array containing the selected SKILL_ID strings in the order they should be considered. Return [] when no Skill is useful. Figure 4: Skill routing prompt. Selects applicable Skills from their manifest before the first generation attempt. ⬇ Produce the one prompt used for the first image-generation attempt. Preserve every requested fact. Apply all selected initial-prompt Skill instructions together as one coherent transformation. Combine compatible guidance according to the original request and return only the resulting prompt. ### Initial-Prompt Skill Instructions SELECTED_SKILL_INSTRUCTIONS ### Original Prompt ORIGINAL_QUERY Return ONLY the final enhanced prompt. Figure 5: Initial prompt rewrite. Combines selected Skill instructions with the original query. ⬇ Analyze the user’s image generation prompt and break it into specific visual requirements. For each requirement, write a question answerable with yes or no. The questions must verify whether the requirement is present in an image. Produce at most 10 independent questions. Prioritize the most important requirements and constraints that an image generator is likely to miss. Respond ONLY with a JSON array of strings. USER PROMPT: ORIGINAL_QUERY Figure 6: Requirement decomposition. Converts the original query into visual yes/no checks. ⬇ Image: <image> Evaluate every verification question below against the provided image. Return ONLY one JSON array with exactly the same number of items and in the same order as the questions. Each item must be an object with an ’answer’ string and a ’passed’ JSON boolean. Use true only when the image clearly satisfies the question; otherwise use false. Do not use markdown or add commentary outside the JSON array. VERIFICATION QUESTIONS: 1. QUESTION_1 2. QUESTION_2 … Figure 7: Batched image verification. Checks an image against all visual requirements in one request. ⬇ Task: Summarize the experience of the current image generation attempt. — CURRENT ATTEMPT — Prompt used: CURRENT_PROMPT Passed requirements: PASSED_REQUIREMENTS Failed requirements: FAILED_REQUIREMENTS Reasoning/Thought before generation: CURRENT_THOUGHT Image: <image> — PREVIOUS EXPERIENCES — PREVIOUS_TASK_LOCAL_EXPERIENCES — ANALYSIS — Based on the image, internal verification results, thought process, and prior attempts from this task, summarize what worked, what failed, and what should be changed in the next attempt. Keep it under 100 words. Do not include an introduction. Figure 8: Task-local experience summary. Records lessons from the current attempt and earlier attempts on the same task. ⬇ Task: Refine the image generation prompt after failed internal checks. ORIGINAL INTENT: ORIGINAL_QUERY — RELEVANT CROSS-TASK INSIGHTS — RELEVANT_INSIGHTS_IF_ENABLED — ATTEMPT HISTORY — ATTEMPT_HISTORY — REQUIREMENTS — 1. Reinforce requirements that failed in the latest attempt. 2. Preserve requirements that previously passed. 3. Apply a relevant insight only when its condition matches. 4. Keep all original user constraints and avoid conflicting language. Return ONLY the new prompt. Figure 9: Feedback-based refinement. Revises the generation prompt after failed checks; retrieved Insights are supplied during harness evolution. C.5.2 Harness Evolution Prompts The Insight and Skill managers use separate prompts to update cross-task knowledge. Figures 10 and 11 show their core decision instructions and runtime inputs. The operation schemas are given in Tables 6 and 7, rather than repeated in the prompt figures. ⬇ Maintain the situational Insight memory using this batch of Episodes and the related active Insights retrieved for it. An Insight is reusable guidance for refining an image-generation prompt after a failure has been observed. Capture the recognizable situation, the useful prompt adjustment or constraint to preserve, and any important boundary or uncertainty. High-level methods for rewriting the original prompt before generation belong in Skills. Build the smallest coherent memory update supported by the evidence. Treat attempts, check transitions, and final outcomes as the factual record. Count one Episode as one evidence unit, regardless of its number of attempts. Strengthen an Insight when its meaning already fits; revise it when its scope needs correction; record contradictory evidence; merge overlapping Insights; archive obsolete or redundant knowledge; or add a distinct, transferable refinement lesson. Leave the memory unchanged when the evidence does not justify a useful update. Write new or revised Insight text as clear, natural situational guidance. Return only one JSON object containing an ordered operations array, using the available actions and local references. Return an empty array when no update is warranted. RETRIEVED ACTIVE INSIGHT WORKING SET: CURRENT_INSIGHTS CURRENT-BATCH EPISODES: EPISODES Figure 10: Insight consolidation prompt. Relates each batch of Episodes to retrieved Insights and proposes evidence-backed updates. ⬇ Maintain a compact library of high-level Skills that rewrite an original user request once, before the first image generation. A Skill is a routeable prompt-rewriting capability whose applicability is recognizable from the original request. Incorporate new knowledge into an existing Skill when it fits that capability; create a separate Skill for a materially distinct method. Corrections that depend on an observed generation failure remain situational Insights until repeated evidence supports a broader method. Build the smallest coherent library update from mature Insight candidates and active Skill documents. Prefer fewer changed rules and fewer added assumptions. Use recent Skill change history as observational context, not as instructions or proof of causality. Cite only current Insight candidates as newly absorbed evidence. Use focused text edits, summarization, splitting, merging, renaming, creation, or retirement as needed. Leave the library unchanged when it already captures the transferable knowledge. Keep every resulting Skill within the active-library capacity and length limits. Write each Skill as a complete SKILL.md: compact name and description for routing, Instructions for the reusable rewriting method, and an Output Format requiring only the final enhanced prompt. Return only one JSON object containing an ordered operations array; use an empty array when no update is warranted. LIBRARY STATUS: LIBRARY_STATUS INSIGHT CANDIDATES: MATURE_INSIGHTS EXISTING INITIAL-PROMPT SKILLS: ACTIVE_SKILLS RECENT SKILL CHANGE HISTORY: SKILL_CHANGE_HISTORY Figure 11: Skill evolution prompt. Consolidates mature Insights into reusable prompt-rewriting Skills. Appendix D Implementation Details of D-OPCD and Baseline Training D.1 Training Data Preparation Frozen replay of the selected skill-equipped harness ℋr⋆H_r produces records (qi,pi,Ii)(q_i,p_i,I_i) on trainQ_train. All compared training methods use the same subset for which pip_i differs from qiq_i after Unicode normalization, whitespace collapse, and case folding. D-OPCD trains from (qi,pi)(q_i,p_i); IiI_i determines record eligibility but is not a loss target. Diffusion-DPO additionally pairs the submitted image IiI_i with a rejected image generated directly from qiq_i using the same seed. Vanilla SFT and D-OPSD train from the corresponding (qi,Ii)(q_i,I_i) records. The per-benchmark counts and filtering yields are reported in Appendix E.2. D.2 D-OPCD Implementation The student receives qiq_i. For the teacher, the text encoder receives qiq_i followed by a blank line, the label Privileged generation context:, and pip_i. The student and teacher use the same frozen text encoder, with maximum sequence lengths of 512 and 1024 tokens, respectively. Starting from Gaussian noise, the student follows a four-point Euler rollout at t∈0,0.1,0.25,0.5t∈\0,0.1,0.25,0.5\, with its final update ending at t=1t=1. The state is detached before each velocity comparison, so gradients update the student at visited states without backpropagating through earlier rollout steps. An EMA copy of the student, with decay 0.99990.9999, supplies the teacher prediction under the additional context. We minimize the mean squared velocity difference across the four states, as defined in Section 3.2; the teacher prediction is stop-gradient. D.3 Baseline Objectives Vanilla SFT. The student receives qiq_i, while the agent-submitted image IiI_i supplies the clean image latent for the standard flow-matching target. Each batch samples a noise level from a logit-normal distribution with logit mean zero and standard deviation one. No privileged prompt enters the model. Diffusion-DPO. Both the chosen image IiI_i and the same-seed rejected image are trained under qiq_i. A frozen copy of the initial generator supplies reference losses, and the chosen–rejected difference in model and reference flow-matching losses forms the preference objective with β=2500β=2500. Two logit-normal noise levels are sampled per pair. D-OPSD. The student receives qiq_i and visits states along its own sampling trajectory. The EMA teacher, updated with decay 0.99990.9999, receives an embedding of (qi,Ii)(q_i,I_i) from Qwen3-VL-4B-Instruct and supplies a detached clean-output endpoint target. This baseline shares SFT’s image-paired records, but uses on-policy student states rather than the image-derived states used by SFT. D.4 Optimization and Checkpoint Selection All methods start from Z-Image-Turbo and train LoRA adapters on its transformer feed-forward projections and attention query, key, value, and output projections. The common settings are rank 64, scale parameter α=128α=128, zero LoRA dropout, 1024-pixel square images, and seed 42. We use AdamW with β1=0.9 _1=0.9, β2=0.999 _2=0.999, zero weight decay, a constant learning rate of 10−410^-4 without warmup, gradient clipping at norm 1, and bfloat16 precision. Training runs for 2000 optimizer steps and saves a checkpoint every 200 steps. Direct evaluation uses eight sampling steps and guidance scale zero. Compute resources. Model training and image-generation inference used NVIDIA A100 GPUs (80 GB each). Training jobs allocated four GPUs; inference used one GPU per generation worker, with multiple workers run in parallel where applicable. Each saved checkpoint is evaluated on holdQ_hold; the benchmark-specific selection metrics are defined in Appendix E. Table 11 lists each benchmark and compared method using skill-equipped harness data, with the reported optimizer step and held-out score. Hold scores use the same benchmark-native metric used for checkpoint evaluation and for the main evaluation tables. Table 11: Generator checkpoints and held-out scores by method and benchmark. All methods use records from ℋr⋆H_r ; scores use GenEval Overall, GenEval2 Soft-TIFA AM, WIScore, and R2I-Score (0–100). Benchmark Step Hold score D-OPCD GenEval 1800 86.31 GenEval2 1600 82.38 WISE 1000 48.67 R2I-Bench 400 49.33 Vanilla SFT GenEval 1000 82.98 GenEval2 1600 82.13 WISE 200 50.67 R2I-Bench 600 51.30 Diffusion-DPO GenEval 200 79.29 GenEval2 1400 79.49 WISE 1800 49.33 R2I-Bench 1200 47.87 D-OPSD GenEval 1200 79.64 GenEval2 1800 75.29 WISE 400 51.33 R2I-Bench 200 49.05 Appendix E Experimental and Evaluation Details E.1 Task Set Allocations Table 12 gives the disjoint task sets used for harness evolution, selection, and final evaluation. GenEval, GenEval2, and WISE use approximately 50/15/35 allocations; R2I-Bench uses a 780-task experimental subset with its supplied split. Table 12: In-distribution task allocations. Ratios give the train/hold/evaluation shares of the task pool actually used. Benchmark Pool |train||Q_train| |hold||Q_hold| |eval||Q_eval| Ratio (%) GenEval 553 276 83 194 49.9/15.0/35.1 GenEval2 800 400 120 280 50/15/35 WISE Verified 1000 500 150 350 50/15/35 R2I-Bench subset 780 480 100 200 61.5/12.8/25.6 The first three splits use seed 42; GenEval2 and WISE preserve earlier task assignments, while R2I-Bench uses its supplied split. E.2 Replay Data Construction and Counts Frozen replay on trainQ_train yields one submitted (qi,pi,Ii)(q_i,p_i,I_i) record per task. Table 13 reports the records retained before and after excluding normalized pi=qip_i=q_i pairs. D-OPCD, Diffusion-DPO, Vanilla SFT, and D-OPSD use the prompt-changed tasks, with method-specific supervision. The replay uses the selected skill-equipped harness version in Table 10. Table 13: Replay-record yields from the selected skill-equipped harness ℋr⋆H_r . The last column gives the prompt-changed tasks used by all four compared training methods. Benchmark Source Records pi=qip_i=q_i Changed GenEval ℋr⋆H_r 276 0 276 GenEval2 ℋr⋆H_r 400 0 400 WISE ℋr⋆H_r 500 10 490 R2I-Bench ℋr⋆H_r 480 88 392 E.3 Context Effectiveness and Diversity Diagnostics Improvement-filtered pairs. The default set retains records from the selected skill-equipped harness when the submitted prompt differs from the original query after normalization. The improvement filter instead starts from all saved records, including unchanged prompts. For each task, we pair the saved image from the harness’s selected attempt with the first image generated directly from qiq_i. We take the requirement checklist iC_i from the selected harness attempt and score both saved images against that same checklist using the harness’s internal verifier. Previously recorded answers are reused when the checklist matches; otherwise, the direct-query image is verified against iC_i. If Vi(I)V_i(I) is the number of passed checks for image I, we retain (qi,pi)(q_i,p_i) iff Vi(Iip)>Vi(Iiq)V_i(I_i^p)>V_i(I_i^q). Missing or invalid verifier responses are excluded. This comparison requires neither new image generation nor external benchmark scores. Since the saved images may use different seeds, the filter selects observed verifier improvements rather than estimating a seed-controlled prompt effect. It selects 43, 264, 319, and 314 pairs on GenEval, GenEval2, WISE, and R2I-Bench, respectively. Three R2I-Bench pairs have pi=qip_i=q_i but satisfy the image-level criterion. Diversity-adapted sampling. When the Skill router selects applicable Skills, the default harness combines their guidance with qiq_i in one initial prompt rewrite. To broaden the contexts produced by the same Skills, the diversity adapter acts at this step, before the first image-generation attempt. It samples one of four expression strategies at random for each task with selected Skills. Local edit preserves the query’s wording and order where possible while adding relevant guidance. Restate rewrites the request in new words. Integrate weaves guidance into a single coherent image description. Separate states the original request first and places guidance in separate sentences. All four instruct the rewriter to preserve the task requirements rather than create variety by changing the requested content. Table 14 gives the exact strategy instructions, and Figure 12 shows the shared rewrite prompt that receives qiq_i, the selected Skill instructions, and the sampled strategy. If no Skill is selected, the first attempt uses qiq_i directly. The Skill library, router, internal verifier, and later refinement steps remain the same, with at most five generation attempts per task. We record the prompt of the agent’s final selected image as pip_i, which need not be its first or last attempted prompt. After excluding normalized pi=qip_i=q_i cases, the adapted replay yields 276, 400, 491, and 381 pairs on GenEval, GenEval2, WISE, and R2I-Bench. All three context sets use the same D-OPCD objective and student/[qi;pi][q_i;p_i] teacher interface. Table 14: Exact expression-strategy instructions inserted into the diversity-adapter rewrite template. Mode Instruction local_edit Keep the original request’s wording and order where possible. Incorporate applicable Skill guidance through local edits or additions. restate Write the resulting prompt afresh in your own words rather than retaining the original sentence structure. Express the same request together with applicable Skill guidance. integrate Weave applicable Skill guidance directly into the description of the request, as one coherent image-generation prompt. separate State the original request first, then express applicable Skill guidance in separate sentences within the same final image-generation prompt. Do not output explanations of your work. ⬇ Produce the prompt for the first image-generation attempt. Use the supplied Skills as guidance for fulfilling the original request. Apply guidance only where it is relevant to that request, and preserve the request’s meaning and requirements. If a Skill conflicts with the original request, the request takes priority. Return only the resulting prompt. Returning the original request unchanged is allowed. ### Initial-Prompt Skill Instructions SELECTED_SKILL_INSTRUCTIONS ### Original Prompt ORIGINAL_QUERY ### Expression strategy STRATEGY_INSTRUCTION This strategy changes expression, not the requested content. It does not authorize dropping applicable guidance or adding new requirements merely to create variety. Return ONLY the final enhanced prompt. Figure 12: Diversity-adapter rewrite template. The strategy instruction is chosen from Table 14; all other text is shared across modes. Paired task effectiveness. For each retained (qi,pi)(q_i,p_i), we generate one image from qiq_i and one from pip_i using the same initial generator, seed, sampling settings, and task. Both images are scored against the original qiq_i and its benchmark annotations. Let Sb(q)S_b(q) and Sb(p)S_b(p) denote the benchmark-native aggregate scores for setting b on exactly its retained tasks, expressed on a 0–100 scale. Table 15 reports Δb=Sb(p)−Sb(q) _b=S_b(p)-S_b(q) in percentage points, 100Δb/Sb(q)100 _b/S_b(q) as a relative percentage, and the percentage of individual tasks whose pip_i image scores strictly above its paired qiq_i image. GenEval uses Overall and GenEval2 uses Soft-TIFA AM. The same-seed external evaluation is distinct from the agent’s internal-verifier criterion used for filtering. Table 15: Effectiveness of the three training-pair settings on their retained train tasks. S⁡(q)S(q) and S⁡(p)S(p) use the original query as the scoring target: GenEval Overall, GenEval2 Soft-TIFA AM, WIScore, or R2I-Score, respectively. Δ=S⁡(p)−S⁡(q) =S(p)-S(q) is in percentage points; Rel. is 100Δ/S⁡(q)100 /S(q), and Win is the percentage of individual pairs with a strictly higher p-image score. The external benchmark scorer evaluates both images in each pair. Benchmark Setting N S⁡(q)S(q) S⁡(p)S(p) Δ Rel. (%) Win (%) GenEval Default 276 73.76 85.71 +11.95 +16.20 19.57 Improvement filter 43 12.50 78.99 +66.49 +531.94 69.77 Diversity adapter 276 73.76 81.84 +8.08 +10.95 16.30 GenEval2 Default 400 77.43 85.74 +8.31 +10.73 70.75 Improvement filter 264 73.37 85.35 +11.98 +16.33 78.79 Diversity adapter 400 77.43 86.76 +9.33 +12.06 72.75 WISE Default 490 43.74 80.00 +36.26 +82.89 38.78 Improvement filter 319 29.92 79.91 +49.99 +167.12 51.72 Diversity adapter 491 43.87 80.67 +36.80 +83.90 40.33 R2I-Bench Default 392 42.58 77.17 +34.59 +81.23 75.00 Improvement filter 314 37.87 76.43 +38.56 +101.83 79.62 Diversity adapter 381 42.48 75.79 +33.31 +78.43 74.54 Context-text diversity. We jointly fit character 3–5-gram TF–IDF features to all three settings within each benchmark. We measure normalized character Vendi score (Friedman and Dieng, 2023) (Vendi divided by sample size), mean within-set pairwise cosine similarity, and corpus Distinct-1 (Li et al., 2016) (unique words divided by all words). Mean prompt length aids interpretation. To reduce the effect of cohort size, Table 16 uses equal-size subsamples, fixed by the smallest setting in each benchmark: 43/264/319/314 for GenEval/GenEval2/WISE/R2I-Bench. For larger settings, it averages 100 subsamples drawn with seed 42. These are text-diversity proxies rather than image-diversity scores. The R2I-Bench improvement-filtered pool includes three unchanged pi=qip_i=q_i pairs; we retain them to match the model-training data. A separate earlier matched-400 GenEval2 audit compared the default SkillOnly contexts with those from the diversity adapter, fitting TF–IDF jointly to those two corpora. It found normalized lexical Vendi scores of 0.7402 and 0.7921, respectively, and Distinct-2 values of 0.3533 and 0.3862. Because this two-setting word-level audit used a different feature space and sample protocol, its numbers are not inserted into Table 16. Table 16: Text diversity of the retained privileged contexts pip_i. Normalized character Vendi and Distinct-1 are higher for more varied contexts; mean pairwise character TF–IDF cosine is lower. Pool N is the number of retained pairs, and Eval n is the equal-size subsample. Length is the mean number of words. Benchmark Setting Pool N Eval n Char Vendi ↑ Char cosine ↓ Distinct-1 ↑ Length GenEval Default 276 43 0.7754 0.1519 0.1768 96.4 Improvement filter 43 43 0.7864 0.1431 0.1767 116.7 Diversity adapter 276 43 0.7841 0.1459 0.1817 83.5 GenEval2 Default 400 264 0.4914 0.1374 0.0483 189.3 Improvement filter 264 264 0.4816 0.1386 0.0476 197.4 Diversity adapter 400 264 0.5664 0.1094 0.0590 111.0 WISE Default 490 319 0.7583 0.0796 0.1680 112.3 Improvement filter 319 319 0.7339 0.0856 0.1592 122.9 Diversity adapter 491 319 0.8133 0.0610 0.1940 84.7 R2I-Bench Default 392 314 0.7668 0.0708 0.1183 143.7 Improvement filter 314 314 0.7582 0.0721 0.1135 153.3 Diversity adapter 381 314 0.7946 0.0632 0.1265 123.5 Character n-grams probe repetition in the prompt wording that the adapter is designed to reduce. A word unigram–bigram audit agrees on GenEval2, WISE, and R2I-Bench, but its GenEval differences are small and mixed: normalized word Vendi is 0.9221 for the default and 0.9205 for the adapter, while Distinct-2 is 0.5469 and 0.5430, respectively. E.4 Evaluation Metrics and Protocols For in-distribution evaluation, each method generates one image per task on evalQ_eval. We report each benchmark by its native metric under its official evaluation protocol. GenEval (Ghosh et al., 2023) uses the Overall score from its detector-based checker; GenEval2 (Kamath et al., 2025) uses prompt-level Soft-TIFA AM; WISE (Niu et al., 2025) uses WiScore under the WISE_Verified binary protocol; and R2I-Bench (Chen et al., 2025) uses R2I-Score from its checklist-based judge. Checkpoint candidates are evaluated on holdQ_hold, with Overall score for GenEval, Soft-TIFA AM for GenEval2, WiScore for WISE, and R2I-Score for R2I-Bench. Reported hold scores use these same evaluation metrics. T2I-CompBench++ (Huang et al., 2025) is evaluated separately on 2,100 prompts from the seven categories in Table 5 of its official validation split, using three images per prompt with seeds 42–44 and reporting scores by category. This OOD set is not holdQ_hold and is used neither for skill evolution nor for model or harness selection. The base generator is evaluated once, and each other row in Table 5 reports one source-specific D-OPCD checkpoint, without averaging across training sources. Appendix F Use of AI Assistants AI assistants were used in three roles. First, the experimental harness uses GPT-5.6 Terra to plan and refine generation prompts, verify images, and consolidate task experience into Insights and Skills (Appendix C). Second, outputs are scored with the benchmark-native automated evaluators described in Appendix E, including the checklist-based judge for R2I-Bench. Third, general-purpose AI writing tools were used to improve wording and readability. They were not used to originate the research questions, choose the experimental design, produce the reported results, or draw the paper’s conclusions. The authors reviewed and approved the experimental design, analyses, results, and manuscript text.

Instant research discovery

Internalizing Agent Experience into Diffusion Model Weights via On-Policy Context Distillation

Intelligence

Summary

Entities (8)

Relation Signals (7)

Abstract

Tags

Links

PDF

Full Text

Paper	Published	Area	Tags	Intel	Citations