PaperlayerView exact revision on arXiv

Source-linked AI summary

GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation

Sixiang Chen, Zhaohu Xing, Tian Ye, Xinyu Geng, Yunlong Lin, Jianyu Lai, Xuanhua He, Fuxiang Zhai, Jialin Gao, Lei Zhu

arXiv:2605.21605v2cs.CV

TL;DR

Open-ended image generation requires coordinating internal knowledge with external evidence, references, and tools beyond one-shot prompting. GenEvolve learns such tool-orchestrated trajectories through teacher-only visual experience distillation, improving agentic generation and final image quality, including a KScore increase from 0.3493 to 0.3663 over Gen-Searcher.

  • Problem

    Open-ended image generation needs agents that coordinate internal generative knowledge with external evidence, references, and tools across complex grounded requests.

  • Method

    GenEvolve models generation as tool-orchestrated visual trajectories and distills best-worst differences from multiple attempts into teacher-only token-level visual experience.

  • Results

    0.3663 KScore, up from 0.3493 for Gen-Searcher, while GenEvolve achieves the best scores on both Knowledge-Anchored and Quality-Anchored tracks.

  • Takeaways & Limitations

    GenEvolve shows that learning complete tool-orchestrated generation trajectories can improve agentic behavior and final image quality over traditional generation.

Abstract

from arXiv · show

Open-ended image generation is no longer a simple prompt-to-image problem. High-quality generation often requires an agent to combine a model's internal generative ability with external resources. As requests become more diverse and demanding, we aim to develop a general image-generation agent that can self-evolve through trajectories and use tools more effectively across varied generation challenges. To this end, we propose GenEvolve, a self-evolving framework based on Tool-Orchestrated Visual Experience Distillation. In GenEvolve, each generation attempt is modeled as a tool-orchestrated trajectory, where the agent gathers evidence, selects references, invokes generation skills, and composes them into a prompt-reference program. Unlike existing agentic generation methods that mainly rely on image-level scalar rewards, GenEvolve compares multiple trajectories for the same request and abstracts best-worst differences into structured visual experience, provided only to a privileged teacher branch. Inspired by on-policy self-distillation, Visual Experience Distillation provides dense token-level supervision, helping the student internalize better search, knowledge activation, reference selection, and prompt construction. We further construct GenEvolve-Data and GenEvolve-Bench. Experiments on public benchmarks and GenEvolve-Bench show substantial gains over strong baselines, achieving state-of-the-art performance among current image-generation frameworks. Our website is as follows: https://ephemeral182.github.io/GenEvolve/

1 Introduction

GenEvolve reframes open-ended image generation as tool-orchestrated trajectory learning, coordinating external evidence, visual references, internal generation knowledge, and prompt-reference programs. Its visual experience distillation, trajectory data, and benchmark support self-evolution beyond image-level reward signals.

  • Motivation: Open-ended image generation requires agentic planning to acquire evidence, select references, activate knowledge, and translate these signals into generator-followable instructions.The introduction emphasizes that fidelity alone is insufficient for complex, grounded, and intent-sensitive requests.
  • GenEvolve Framework: GenEvolve models each generation attempt as a visual trajectory that gathers textual evidence, selects references, invokes callable knowledge, and synthesizes z = (g, R).Here, g is a targeted prompt and R is a small set of selected reference images.
  • Data and Benchmark: GenEvolve-Data provides tool-orchestrated trajectories and filtered ground-truth image cases for acquiring evidence, activating knowledge, constructing programs, and making visual feedback meaningful.The data extends beyond ordinary prompt-rewriting corpora.
  • Data and Benchmark: GenEvolve-Bench evaluates final image quality across knowledge- and quality-oriented dimensions, while the agent coordinates search, retrieval, generation knowledge, synthesis, image generation, and experience internalization.The stated contributions position the framework as a general image-generation agent trained through coordinated tool use.
  • Visual Experience Distillation: GenEvolve compares multiple trajectories for one request and abstracts best-worst differences into structured visual experience instead of relying only on image-level scalar rewards.This approach aims to explain which decisions caused improvement and supports on-policy self-distillation.

2 Related Work

Related work spans increasingly capable image-generation models, agentic systems that add planning and tools, and on-policy distillation methods for training agents. GenEvolve builds on these directions by coordinating external tools with internal generation knowledge and distilling visual experience along trajectories.

  • Image generation models: Diffusion and latent diffusion models established high-fidelity prompt-conditioned synthesis, while newer diffusion transformers improve scalability, text understanding, and generation quality.The passage names DiT, PixArt-α, Stable Diffusion 3, FLUX, Hunyuan-DiT, and Nano Banana Pro as successors.
  • Agentic image generation: Agentic image-generation systems augment image models with planning, retrieval, tool use, judging, or refinement.GenAgent supports multi-turn reasoning, tool invocation, judgment, and reflection; Mind-Brush, Gen-Searcher, and ORIG emphasize research, search, or retrieval-augmented generation.
  • Agentic image generation: Existing agentic systems demonstrate the value of search, tools, memory, and refinement but often emphasize one component or wrap a generator with an external workflow.This motivates a more integrated generation process.
  • Agentic image generation: GenEvolve coordinates external tools and internal generation knowledge along visual trajectories and uses visual experience distillation to improve their coupling.The framework is described as an open image-generation agent inspired by tighter integration of reasoning, knowledge, grounding, and visual synthesis.
  • On-policy distillation: On-policy distillation methods supervise on-policy generations, distill in-context knowledge into parameters, convert rich feedback into dense signals, or summarize trajectories into training-only skills.The passage identifies OPSD, OPCD, SDPO, and Skill-SD as representative variants.

3 Tool-Orchestrated Visual Trajectory Formulation

GenEvolve models each image-generation attempt as an observable, trainable tool-orchestrated visual trajectory. The agent learns to coordinate external information, visual references, internal generation knowledge, and prompt-reference synthesis, while privileged self-distillation adds structured differences between higher- and lower-reward trajectories.

  • Trajectory formulation: Each trajectory lets the agent decide when to acquire information, trust visual references, activate internal generation knowledge, and synthesize a prompt-reference program.This formulation covers both external tool use and internal generation decisions.
  • Trajectory formulation: At each turn, the agent samples either a tool call or final answer from its history-conditioned policy and receives the corresponding observation.The interaction history contains the request and preceding action-observation pairs.
  • Trajectory formulation: The final answer is a prompt-reference program z = (g, R), pairing a targeted generation prompt with a small set of selected reference images.A reference-conditioned generator then renders the result from this program.
  • Learning signal and scope: Unlike prior systems that optimize loosely coupled interfaces, GenEvolve treats tool use, knowledge activation, reference selection, and prompt synthesis as trajectory-level learning decisions.For the same request, it compares higher- and lower-reward trajectories and gives their structured visual differences only to a privileged teacher branch.

4 GenEvolve-Data and GenEvolve-Bench

GenEvolve-Data is a complete tool-orchestrated generation pipeline that converts diverse prompts into audited teacher trajectories, filtered ground-truth image cases, and separate training, self-evolution, and evaluation views. GenEvolve-Bench is its held-out evaluation split for testing evidence use, reference selection, and quality-aware generation under KScore.

  • GenEvolve-Data: GenEvolve-Data converts diverse prompts into tool-using teacher solutions, audits them with VLM filters, renders ground-truth image cases, and creates separate dataset splits.The pipeline supports supervised cold start, self-evolution, and held-out evaluation.
  • Prompt pool: The prompt pool contains Knowledge-Anchored requests requiring external grounding and Quality-Anchored requests emphasizing requirements such as text layout, composition, counting, anatomy, and materials.Prompts are generated from structured recipes specifying task family, missing evidence, visual anchor, dominant requirement, and difficulty.
  • Teacher trajectories: Teacher trajectories use Seed2.0 and Gemini 3 Pro in real multi-turn tool loops to search, retrieve references, activate generation knowledge, and synthesize prompt-reference programs.Tool order depends on the request, including factual lookup for knowledge-heavy cases.
  • Trajectory filtering: Programmatic checks and a VLM judge remove invalid trajectories and assess reference support, evidence use, and integration of required constraints.Checks target incomplete loops, invalid references, URL or ID leakage, missing ordinal wording, and underspecified final programs.
  • GT images and splits: Surviving cases provide an SFT view without ground-truth images and a visual-feedback view with requests, ground-truth images, and metadata, preventing teacher-output copying.The two views support supervised cold start, self-evolution, and benchmark evaluation.
  • GenEvolve-Bench: GenEvolve-Bench is the held-out split evaluated with KScore on Knowledge-Anchored and Quality-Anchored subsets, testing evidence, reference, and quality-control integration.It is designed to distinguish agentic generation from generic text-to-image prompting.

5 Method: GenEvolve

GenEvolve trains an image-generation agent through multi-turn tool-orchestrated visual trajectories that produce executable prompt-reference programs. It self-evolves by combining supervised cold-start training, dual-reward GRPO, and privileged visual experience distillation from better and worse generated trajectories.

  • Trajectory formulation: Each request is solved through a multi-turn trajectory of tool calls, observations, and an executable prompt-reference generation program.The agent follows a ReAct-style interface rather than performing a single prompt rewrite.
  • Tool orchestration: The action space combines textual search, visual-reference retrieval, and callable internal generation skills for tasks such as layout, anatomy, counting, and text rendering.The skills also cover attribute binding, material consistency, aesthetics, and creative transformation.
  • Training pipeline: GenEvolve-Data cold-starts the base MLLM to decide when to call tools, select references, activate knowledge, and output a valid program z = (g, R).After supervised trajectory training, GRPO and Visual Experience Distillation further optimize the initialized policy with generated-image feedback.
  • Prompt-reference program: The final program contains a targeted instruction g and an ordered reference set R, with ordinal phrases binding selected images to user constraints, retrieved facts, knowledge, and failure-avoidance experience.This program is passed to a reference-conditioned generator.
  • Trajectory optimization: GenEvolve uses dual rewards: Rimg judges faithfulness, visual correctness, text accuracy, and aesthetics, while Rtext measures program sufficiency.Group-relative policy optimization compares K sampled trajectories using the mixed reward.
  • Visual Experience Distillation: Visual Experience Distillation extracts best-worst trajectory differences into structured lessons and gives them only to a privileged teacher for dense token-level guidance during training.Unlike methods distilling privileged answers or reasoning traces, GenEvolve distills generated visual experience to improve tool use, knowledge activation, reference selection, and prompt-reference synthesis.

6 Experiments

GenEvolve is evaluated on held-out GenEvolve-Bench prompts using KScore and four rubric dimensions, with comparisons spanning direct generators, agent frameworks, and downstream-generator pairings. It also demonstrates external generalization on WISE and incremental gains across training stages.

  • Evaluation setup: GenEvolve-Bench compares direct image generators, recent agent frameworks, and GenEvolve paired with open and strong downstream generators under shared prompts and judge rubrics.Reported metrics include KScore, faithfulness, visual correctness, text accuracy, and aesthetics.
  • Main comparisons: 0.5298 KScore is achieved by Nano Banana Pro, compared with 0.2987 for Qwen-Image, while demanding grounded cases benefit from agent-side orchestration.The benchmark contains Knowledge-Anchored and Quality-Anchored prompts requiring factual grounding, reference use, and quality-sensitive prompt-reference construction.
  • External generalization on WISE: 0.82 WiScore makes GenEvolve the top overall method on WISE, surpassing GPT-4o at 0.80 and agentic baselines including GenAgent at 0.72, Gen-Searcher-8B at 0.77, and Mind-Brush at 0.78.The strongest category results are 0.83 for chemistry and 0.83 for biology.
  • Ablation study: 0.3548 KScore is reached by adding GRPO without visual experience, after direct Qwen-Image at 0.2987, an untuned workflow at 0.3317, and SFT cold start at 0.3480.The ablation attributes the untuned-workflow gain to tool-interface use and reference-conditioned generation, and the SFT gain to curated trajectories.

7 Conclusion

GenEvolve is presented as a self-evolving image-generation framework based on Tool-Orchestrated Visual Experience Distillation. It models generation as a visual trajectory that coordinates external evidence, references, internal knowledge, prompt-reference construction, generation, and feedback, then extracts structured experience by comparing best and worst trajectories.

  • Conclusion: GenEvolve is a self-evolving framework for image-generation agents based on Tool-Orchestrated Visual Experience Distillation.The framework is presented as an alternative to a single prompt-to-image call.
  • Conclusion: Generation is formulated as a visual trajectory coordinating external evidence, visual references, internal generation knowledge, prompt-reference construction, generation, and feedback.
  • Conclusion: GenEvolve compares best-worst trajectories to extract structured visual experience for distillation in the teacher-only branch.The supplied passage ends mid-word after describing this teacher-only distillation process.

A GenEvolve-Data Construction … A.6 Coverage and Construction Statistics

GenEvolve-Data is built as complete, tool-orchestrated visual generation problems spanning supervised cold start, self-evolution, and held-out evaluation. Its construction combines recipe-controlled prompts, filtered teacher trajectories and GT images, complementary exports, and stratified coverage.

  • GenEvolve-Data Construction: GenEvolve-Data supports supervised cold start, self-evolution, and held-out evaluation through complete visual generation problems requiring evidence, references, knowledge, and prompt-program synthesis.Examples are constructed as full generation tasks rather than ordinary prompt-rewriting instances.
  • A.1 Prompt Pool Recipes: The prompt pool contains 19,990 valid prompts: 11,999 Knowledge-Anchored and 7,991 Quality-Anchored, averaging about 65 words.Difficulty counts are 13,333 hard, 6,654 medium, and 3 easy prompts.
  • A.2 Teacher Trajectory Generation: Each validated prompt becomes a real multi-turn teacher trajectory that alternates reasoning, tool calls, observations, and final prompt-reference programs.Accepted trajectories require meaningful tool use and image search, while allowing diverse tool-order patterns and optional generation-knowledge calls.
  • A.3 Filtering Rubric: Filtering combines hard protocol checks with VLM judgment across prompt suitability, reference grounding, process quality, skill integration, faithfulness, and training value.Hard checks remove incomplete, malformed, unsafe, or protocol-invalid trajectories before semantic review.
  • A.3 Filtering Rubric: 13,379 of 19,320 structurally valid trajectories were kept (69.2%).Average skill integration was 4.70/5.
  • A.4 GT Image Generation and Filtering: 4,321 successful GT images were generated from 4,379 attempts (98.7%), with 3,175 retained after filtering (73.5%).Nano Banana Pro generated the images, which were filtered for prompt compliance, reference utilization, coherence, and quality.
  • A.5 Supervised and Self-Evolution Export: The self-evolution view uses 3,175 filtered GT-image cases, split into a 2,575-case training pool and about 600 evaluation cases.The training pool contains 2,446 optimization cases and 129 internal validation cases; teacher trajectories and final programs are removed.
  • A.6 Coverage and Construction Statistics: The dataset organizes Knowledge-Anchored and Quality-Anchored tracks across eight diagnostic categories each, with a held-out benchmark split having no exact overlap with self-evolution training.The two-track hierarchy supports coverage control, split stratification, and benchmark analysis.

B Additional Method Details · B.1 Tool-Orchestrated Rollout Protocol · B.2 Prompt-Reference Program Schema

GenEvolve’s supplementary method details define an auditable tool-orchestrated rollout protocol and a portable prompt-reference generation-program schema. The protocol permits adaptive tool use, while the schema orders local references and records queried generation knowledge for execution and training.

  • B Additional Method Details: The supplementary section specifies implementation details for rollout, prompt-reference programs, experience memory, retrieval, GRPO loss, and experience-conditioned self-distillation.These details complement the main method while keeping the core paper concise.
  • B.1 Tool-Orchestrated Rollout Protocol: At inference, each assistant turn emits exactly one tool call or terminates with a parseable final answer.The environment executes the tool, appends its observation, and resumes the agent.
  • B.1 Tool-Orchestrated Rollout Protocol: The rollout treats invalid tool names, invalid knowledge names, missing references, raw URL leakage, and unparseable final JSON as trajectory errors.These constraints reinforce the protocol’s auditable, parseable execution format.
  • B.1 Tool-Orchestrated Rollout Protocol: Tool use is adaptive rather than fixed: prompts may search before image_search, skip text search, or call several generation-knowledge tools.Knowledge-Anchored prompts often search before image_search, Quality-Anchored prompts may skip text search, and complex prompts may call several generation-knowledge tools.
  • B.2 Prompt-Reference Program Schema: The final executable object is z = (g, R), combining a natural-language instruction g with an ordered list R of local reference images.The instruction is consumed by a reference-conditioned generator.
  • B.2 Prompt-Reference Program Schema: Instructions must identify selected references ordinally, such as “the first reference image,” instead of using raw image IDs or URLs.This makes the program independent of transient retrieval IDs and compatible with generators receiving images as ordered inputs.
  • B.2 Prompt-Reference Program Schema: The schema records generation prompts, reference images with roles, selected skills, and a rationale explaining the evidence, references, and knowledge.The example includes identity/shape and style-or-layout reference roles, plus spatial_layout and text_rendering skills.
  • B.2 Prompt-Reference Program Schema: Selected skills are not hidden generator controls; they document queried internal knowledge and should be reflected in g.During training, the field supports tool-call supervision, diagnostic analysis, and visual-experience slot construction.

B.3 Visual Experience Extraction — Real Training Cases … B.6 GRPO Rollout Loss

GenEvolve distills best-worst rollout differences into prompt-keyed visual experience that guides privileged-teacher training. The framework organizes lessons across search, skill activation, reference selection, prompt construction, and failure avoidance, while GRPO optimizes assistant decisions using group-relative rewards.

  • B.3 Visual Experience Extraction — Real Training Cases: Best-worst trajectory comparisons expose failures in search formulation, skill activation, reference selection, prompt synthesis, and failure avoidance.Three cases show noisy factual grounding, skipped generation knowledge, and vague spatial prompting as distinct failure modes.
  • B.3 Visual Experience Extraction — Real Training Cases: Case 1 shows that adding “flag” to the search query causes noise, incorrect entity identification, downstream Chinese flag grounding, and text reward 0.0.The extracted lesson is to search for the entity first and deduce the required attribute afterward.
  • B.4 Prompt-Keyed Visual Experience Memory: The pipeline stores high-gap comparisons in five prompt-keyed dimensions and trims each capacity-limited slot by reward gap and recency.Comparisons caused only by missing references or protocol failures are ignored.
  • S3 REFERENCE SELECTION: Reference-selection experience improves filtering, deduplication, and ordinal reference binding while identifying useful, redundant, or harmful candidates and recurring low-reward patterns.These lessons are stored with intended reference roles and similar-request failure patterns.
  • S5 FAILURE AVOIDANCE: Failure-avoidance experience adds guards for text, counting, material, anatomy, grounding, and style failures.Case evidence shows that missing spatial anchors can cause buildings to merge and signs to attach incorrectly.
  • B.5 Source-Prompt Bundle Retrieval: Teacher context is retrieved by source-prompt similarity and bundles all five slots from the same source prompt, avoiding unrelated lessons.If no matching entry exists, no teacher context is produced and the SDL term is skipped for that row.
  • B.6 GRPO Rollout Loss: For each prompt, GRPO samples K complete visual rollouts containing tool calls, final answers, selected references, generated images, and scalar rewards.The policy loss applies only to assistant tokens, while the same final reward supplies the group advantage for tool decisions and prompt-reference construction.

B.7 Experience-Conditioned SDL Contexts … C.2 Self-Evolution and SDL Configuration

Experience-conditioned SDL gives a privileged teacher the retrieved visual experience bundle while the student retains the normal inference context, supplying dense token-level guidance on-policy. The training pipeline cold-starts the agent with supervised trajectories, then self-evolves by mining best-worst rollout pairs and applying tightly clipped SDL guidance.

  • B.7 Experience-Conditioned SDL Contexts: The teacher inserts a retrieved source-prompt experience bundle into the system prompt, while the student uses the normal inference context and remains the returned policy.Teacher and student share model weights; the teacher view is detached and privileged only during training, so inference requires no dynamic experience slots.
  • B.7 Experience-Conditioned SDL Contexts: SDL reuses on-policy responses from the group-relative rollout loss, adding dense token-level guidance without a separate offline imitation dataset.Its teacher conditioning object is a prompt-keyed visual experience bundle extracted from generated-image best-worst trajectory pairs, with fixed SDL coefficient and clipping constants across ablations.
  • B.7 Experience-Conditioned SDL Contexts: Together, these opposing and supportive effects reroute divergent decisions while sharpening correct ones, making SDL content-specific and complementary to trajectory-level GRPO reward.The observed guidance tracks factual identity, reference-based counting, and spatial composition rather than acting as a generic regulariser.
  • B.7 Experience-Conditioned SDL Contexts: Teacher guidance opposes less effective lexical choices by redirecting planning, skill activation, and grounding tokens toward experience-recommended search, layout, reference, and factual choices.Examples include “let” →“call”, “shape” →“layout”, and “gen” →“reference”.
  • B.7 Experience-Conditioned SDL Contexts: The teacher also boosts already-correct but uncertain student tokens, including spatial, count, and search decisions tied to the case-specific experience bundle.At T0, probabilities increase from 0.527→0.961 for ‘spatial’, 0.637→0.793 for ‘count’, and 0.622→0.848 for ‘search’.
  • C.1 Supervised Stage Configuration: The supervised stage cold-starts Qwen3-VL-8B-Instruct into a tool-orchestrated image-generation agent using supervised GenEvolve-Data trajectories.Only assistant-side reasoning, tool calls, and final prompt-reference programs are optimized; user prompts and tool observations are masked, while the visual encoder stays fixed.
  • C.2 Self-Evolution and SDL Configuration: Self-evolution samples multiple on-policy rollouts per prompt, renders each prompt-reference program, scores images and programs with image and text judges, and mines best-worst pairs with sufficient reward gaps.These pairs are converted into prompt-keyed visual experience for subsequent training.
  • C.2 Self-Evolution and SDL Configuration: SDL_TOP_K_FRAC=0.1 and ρmax = 2 isolate the few decision tokens where the experience-conditioned teacher disagrees most strongly with the student.The configuration uses an importance-ratio cap of 2.

C.3 Self-Evolution Training Dynamics … D.2 Reward Rubric

GenEvolve’s self-evolution steadily improves reward while reducing SDL loss, as the student absorbs retrieved strategic guidance. GenEvolve-Bench evaluates grounded image quality across complementary prompt tracks using KScore, while a separate text judge supervises training programs.

  • D Evaluation Details: Training combines image and text rewards as R = 0.5 Rimg + 0.5 Rtext, with the text judge scoring prompt-reference programs on five bins.The image reward uses KScore, while program sufficiency is evaluated without showing the generated image.
  • C.3 Self-Evolution Training Dynamics: Mean reward rises steadily despite prompt-dependent variance, indicating progressively higher-quality tool-orchestrated trajectories and prompt-reference programs.The trend reflects improved performance under both image-side KScore and text-side program-sufficiency judgments.
  • C.3 Self-Evolution Training Dynamics: SDL loss decreases without reaching zero because the teacher sees retrieved experience while the student operates under plain inference context.SDL measures reverse-KL divergence on the same on-policy tokens and reflects absorption of privileged decision guidance.
  • D.1 GenEvolve-Bench Categories: GenEvolve-Bench evaluates final images through Knowledge-Anchored and Quality-Anchored tracks covering grounded visual facts and visible generation requirements.Knowledge-Anchored cases emphasize externally grounded entities and facts, whereas Quality-Anchored cases stress text, layout, anatomy, attributes, counting, and physical consistency.
  • D.2 Reward Rubric: The visual judge compares each generated image with the request and fixed ground-truth image using the same four-dimensional KScore rubric for raw and agentic systems.This shared protocol avoids a GenEvolve-specific evaluator and follows KnowGen-style evaluation adopted by recent agentic image-generation work.
  • D.2 Reward Rubric: KScore weights visual correctness and text accuracy at 0.4 each, while faithfulness and aesthetics receive 0.1 each.If readable text is not required, text accuracy is marked unavailable and the score is renormalized over the other dimensions.
  • D.2 Reward Rubric: The text-side reward is used only during GRPO+SDL training, while reported main-table results contain only KScore for direct system comparability.The text judge assesses grounded facts, ordinal reference bindings, activated generation knowledge, and executable constraints in the final program.

D.3 External WISE Evaluation Protocol

The external WISE evaluation uses the original six-category WISE release and applies the same rollout interface as GenEvolve-Bench to generate prompt-reference programs from WISE prompts. Its results are reported in the main text and visualized alongside the internal benchmark.

  • Evaluation protocol: The protocol evaluates the original WISE release across six category groups: culture, time, space, biology, physics, and chemistry.The evaluation uses the original WISE release [26].
  • Evaluation protocol: The agent receives only the WISE prompt and produces a prompt-reference program through the same rollout interface used for GenEvolve-Bench.This matches the rollout interface used for GenEvolve-Bench.
  • Results reporting: External WISE results are reported in the main text in Section 6, Table 2, and visualized alongside the internal benchmark in Figure 1.The appendix provides protocol details for these externally evaluated results.

E Additional Qualitative Results · F Prompt and Template Details · F.4 Representative Implementation Prompt Excerpts

GenEvolve demonstrates generality across downstream generators and diverse generation skills, while its implementation templates formalize tool orchestration, contrastive experience extraction, strict output schemas, and retrieval-gated teacher guidance.

  • E Additional Qualitative Results: The same trained agent policy produces qualitative results with Nano Banana Pro and Qwen-Image-Edit, differing only in the downstream reference-conditioned generator.Held-out examples span spatial layout, text rendering, quantity counting, attribute binding, anatomy and pose, creative drawing, physical material consistency, and aesthetic drawing.
  • E Additional Qualitative Results: Across held-out examples, the agent autonomously retrieves evidence, selects references, activates skills, and composes prompt-reference programs for eight callable generation skills.The listed skills cover spatial layout, text rendering, quantity counting, attribute binding, anatomy and pose, creative drawing, physical material consistency, and aesthetic drawing.
  • F.1 Agent Final Answer Template: The final agent answer is one parseable JSON object whose ordinal image references resolve through an ascending img_id-sorted reference_images list.The gen_prompt uses phrases such as “the first reference image” and excludes IMG_### identifiers and raw URLs.
  • F.2 Judge Output Template: The reward judge returns scalar subscores and diagnostics, including faithfulness, visual correctness, text accuracy, aesthetics, overall reward, failure tags, and per-skill status.Diagnostics support analysis and experience construction rather than being directly optimized as independent rewards.
  • F.3 Experience Bundle Template: Each best-vs-worst comparison becomes one compact experience bundle with a retrieval_key and decision_guidance containing a decision focus and six imperative guidance lists.The retrieval key combines a trigger with a source-prompt summary, while guidance covers tool planning, search, skill routing, reference selection, prompt programs, and failure guards.
  • F.4 Representative Implementation Prompt Excerpts: Decision guidance prioritizes observed best-worst differences, falls back to best-trajectory behavior with “Standard:” when categories lack contrast, and otherwise emits empty lists.Bullets use imperative action-space language and specify concrete fields for tools, queries, skills, references, prompt binding, and failure prevention.
  • F.4 Representative Implementation Prompt Excerpts: The representative summarizer prompt compares best and worst trajectories, their generation programs, called skills, rewards, and judge diagnostics to extract action-level decision guidance.The guidance targets tool ordering, search-query decomposition, skill routing, reference selection, and the final prompt-reference program rather than a case report.
  • F.4 Representative Implementation Prompt Excerpts: The teacher-side experience block is appended only when retrieval cosine similarity exceeds EXPERIENCE_MIN_RETRIEVAL_SIM (0.84); otherwise, the student receives plain context without an SDL signal.When retrieved, the guide is treated as the preferred strategy for the current task rather than as a generic past-experience reference.
Loading 2605.21605v2…