acceptodds
Under review as a conference paper at ICLR 2027

EvoSkillGen: Hierarchical Evolution of Compositional Skills for Agentic Image Generation

11 views

Abstract

Image generators can synthesize high-fidelity images, but complex requests involving external knowledge, compositional and spatial constraints, or targeted edits with content preservation often require more than one-shot synthesis. Agentic image generation addresses such requests through retrieval, reasoning, and planning. Yet tasks demand different operations and information flows; selecting predefined workflows alone does not turn execution experience into improved reusable procedures. We propose **EvoSkillGen**, a self-evolving framework for text-to-image generation and image editing. It represents workflows as executable Skills composed of meta-operations with explicit instructions and artifact dependencies. Skill Adaptation (L1) selects task-specific workflows and constructs temporary Skills during evolution; Skill-Bank Evolution (L2) checks generated images against task requirements and uses execution traces to revise Skills and distill temporary workflows. The quality of generated images guides which candidate changes are retained in the Skill bank. We also introduce GenAgentEval, an evaluation suite built from six existing benchmarks to assess final-image quality, workflow-time overhead, and execution-trace diagnostics. Averaged over both judges, EvoSkillGen improves macro-averaged image-quality scores over the strongest baselines by 0.9–1.8 points for text-to-image and 2.3–2.8 for editing across agent configurations. Among agent methods, its workflow-time overhead is second-lowest for text-to-image and lowest for editing.

Primary area:
generative models
Keywords:
Agentic Image Generation, Self-Evolving Agents

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_9rqoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  1. Method. The paper proposes EvoSkillGen, a self-evolving framework for agentic text-to-image generation and image editing, which represents workflows as executable Skills composed of meta-operations.
  2. Two-level learning. (1) L1 selects or creates a Skill for each task. (2) L2 uses checklist feedback and execution traces to revise, distill, and consolidate Skills, retaining updates only if they win pairwise image comparisons.
  3. Evaluation. The paper introduces GenAgentEval, built from six existing benchmarks, and reports consistent gains over agentic baselines with relatively low workflow-time overhead.
Critical strengths
  1. Interpretable workflow representation. Skills are explicit, reusable, and human-readable, and they evolve without updating model weights.
  2. Consistent improvements. EvoSkillGen achieves the best average results across both agents, both judges, and both modalities, and the ablations confirm the contribution of L2.
Critical weaknesses
  1. Insufficient ablations. (1) The contribution of the Prompt Compiler is not isolated. (2) The selective acceptance mechanism in L2 is also not ablated. It remains unclear how the method would perform if proposed updates were accepted without pairwise comparison.
  2. Incomplete cost analysis. (1) The reported workflow-time overhead excludes image generation, which does not reflect the actual cost of each method. The end-to-end time to obtain the final image should be reported instead. (2) The computational cost, such as the number of API calls and token consumption, is not reported.
  3. Missing intermediate baselines. The current baselines are either Direct, with no agent overhead, or full agentic pipelines. Simpler alternatives in between are not considered, for example: (1) single-call prompt rewriting with the same agent, with or without thinking mode, before passing the prompt to Seedream-4; and (2) direct generation followed only by RefineLoop, which iteratively checks the generated image against the request, produces a corrected candidate, and keeps it only if it wins a pairwise comparison, for up to five rounds, without using any Skill.
Questions for authors

See weaknesses 1-3.

5 comments
AuthorsOpenReview ↗

Author Response (1/2)

W1. Prompt Compiler and selective acceptance

Thank you for these suggestions. Tables R7–R8 equally weight three benchmarks and both Gemini/GPT judges, with equal Basic/UGE weights within ImgEdit. Strong/Weak denote GPT-5.5/Doubao-Seed-2.0-pro.

Prompt Compiler. With fixed instructions, the agent LLM extracts necessary additions from preparation-node outputs (retrieved information, layouts and sketches) and selects references, retaining the original request and source images. The bypass fixes the bank and node outputs, appends text artifacts to the unchanged request in a fixed order, and uses the same image interface for source images and eligible references, with unchanged generator settings. Removing Compiler lowers Strong T2I/editing by 0.8/0.5 points and Weak by 1.8/2.3: both benefit, especially Weak.

Selective acceptance. We compare Full's pairwise acceptance with two alternatives:

  • Without bank-level image comparison: remove comparisons on affected acceptance tasks and, when none is affected, on evolution tasks from source Skills' creation batches. Revisions enter after replaying the current-batch tasks that used the old Skill and passing all original checklist items. New/merged Skills enter after operation/dependency checks. Above capacity, prune the less-used Skill in the most similar-capability pair.
  • Separate rubric scoring: the judge sees each image separately with its task inputs. Score original checklist items 1 if met, 0 if unmet/unknown; average them and compare image scores for task wins/ties/losses. Full's judge, comparison-task selection, acceptance, merging and pruning rules remain unchanged.

All three start from Direct-only banks with the same evolution tasks, order and proposal limits, retaining revision replay and task-local RefineLoop. Full leads in all four settings, by 3.8–6.1 points over removing bank-level image comparison and 0.4–1.9 over rubric scoring.

Table R7. Prompt Compiler and selective-acceptance ablations.

VariantStrong T2IWeak T2IStrong EditWeak Edit
Full89.286.172.469.3
w/o Prompt Compiler88.484.371.967.0
L2 acceptance without bank-level image comparison84.182.366.365.2
Separate rubric scoring in L288.885.571.267.4

Full scores are from Table 3.

W3. Intermediate baselines

R8 uses Full's test tasks, generator and scoring protocol. Both agents used thinking during Full's Skill-bank evolution in the main experiments. Direct passes original inputs to Seedream-4 (Direct-Edit for editing). L1 only creates/executes a temporary Skill per request without bank updates. Full learns a bank through L2, freezes it for testing and executes applicable Skills, using Direct when none applies.

Table R8. Intermediate baselines.

MethodStrong T2IWeak T2IStrong EditWeak Edit
Direct81.681.663.063.0
Single-call prompt rewriting (thinking off)83.181.864.363.1
Single-call prompt rewriting (thinking on)84.283.365.164.5
Direct + RefineLoop84.883.567.165.0
L1 only83.783.165.964.2
Full89.286.172.469.3

Direct averages Image scores in Tables 1–2; L1 only and Full are from Table 3.

Single-call rewriting. The same agent receives the original request and editing source images, rewrites the prompt in one LLM call, then invokes Seedream-4 without Skill routing, retrieval or bank updates. Within each agent, off/on share instructions and generation settings (1K, adaptive aspect ratios), differing only in thinking:

  • GPT-5.5: off uses reasoning_effort="none"; on uses reasoning_effort="medium".
  • Doubao: off uses thinking={"type":"disabled"}; on uses thinking={"type":"enabled"} with reasoning_effort="medium".

Both use max_completion_tokens=8192 for reasoning plus answer tokens; temperature is unset. Thinking improves rewriting by 0.8–1.5 points; Full remains 2.8–7.3 points ahead.

Direct + RefineLoop. Check and correct Direct's image against the request for at most five rounds, without creating, retrieving or evolving Skills. Each round generates at most one candidate, kept only if it wins anonymous comparisons with the current image in both presentation orders. Editing must also preserve required source content at least as well as the current image. Correction and stopping follow Full's RefineLoop settings. This improves over Direct by 1.9–4.1 points; Full remains 2.6–5.3 points ahead.

AuthorsOpenReview ↗

Author Response (2/2)

W2. Complete computational cost

Thank you for highlighting the need for end-to-end latency and API usage measurements.

Tables R14 (T2I) and R15 (editing) add end-to-end time, including image generation or editing, together with API and token usage. The measurements use the main-experiment settings: 600 test tasks per modality (200 per benchmark), GPT-5.5 (Strong) or Doubao-Seed-2.0-pro (Weak), and Seedream-4 at 1K with adaptive aspect ratios. Each method retains its original workflow and settings; EvoSkillGen uses its final frozen banks. The original Time column measures workflow overhead, excluding image synthesis and independent evaluation.

1. End-to-end time

Time runs from method invocation to final-image delivery, including task analysis, routing, preparation, generation/editing, internal checks, selection and retries where used. We average per-request wall-clock time over all 600 tasks. Skill-bank evolution costs are reported in R12; independent test-image scoring is excluded from these measurements.

In Table R14, EvoSkillGen takes 52.3/62.3 s (Strong/Weak), second-fastest among agent methods in both settings, after GenRouter. In Table R15, it is fastest among agent methods with both agents (90.7/86.2 s), reducing time by 44.8%/27.3% relative to Mind-Brush and 67.6%/58.8% relative to RS-Gen.

2. API calls and token usage

Tables R14–R15 sum external calls and LLM usage per request, including internal checks and retries, then average over the same 600 tasks. Image-generation/editing calls are listed separately; other API calls include LLM, search and external tool requests, excluding local functions. Tokens are summed from API usage records, with reasoning tokens included in output. Total tokens are input plus output. Independent benchmark scoring is excluded.

EvoSkillGen uses fewer image calls than RS-Gen: 1.52/1.53 versus 2.68/2.67 for T2I, and 1.42/1.47 versus 4.27/3.30 for editing. Other API calls and total LLM tokens are also lower in all four settings. Total token usage decreases by 57.0%/57.4% for T2I and 86.5%/72.5% for editing (Strong/Weak).

Table R14. T2I time and API usage per request; cells Strong / Weak.

MethodEnd-to-end time (s)Other API callsImage-generation API callsLLM input tokensLLM output tokensTotal LLM tokens
Direct5.5 / 5.50.00 / 0.001.00 / 1.000 / 00 / 00 / 0
GenRouter38.3 / 46.44.87 / 4.811.00 / 1.002,105 / 2,182641 / 4442,746 / 2,626
GEMS262.1 / 220.132.85 / 44.352.46 / 3.7935,017 / 55,0481,918 / 88136,935 / 55,929
Mind-Brush96.8 / 84.62.66 / 2.091.00 / 1.004,109 / 3,9791,340 / 1,1525,449 / 5,131
RS-Gen120.5 / 123.68.31 / 7.952.68 / 2.678,849 / 10,7362,307 / 2,26111,156 / 12,997
EvoSkillGen52.3 / 62.34.76 / 6.591.52 / 1.533,125 / 4,2551,669 / 1,2874,794 / 5,542

WISE, DPG-Bench and SpatialGenEval: 200 test tasks each. All columns use the same 600 tasks and report per-request means.

Table R15. Editing time and API usage per request; cells Strong / Weak.

MethodEnd-to-end time (s)Other API callsImage-editing API callsLLM input tokensLLM output tokensTotal LLM tokens
Direct-Edit11.7 / 11.70.00 / 0.001.00 / 1.000 / 00 / 00 / 0
Mind-Brush164.3 / 118.54.84 / 3.161.00 / 1.006,421 / 8,8131,197 / 1,2147,618 / 10,027
RS-Gen279.8 / 209.211.69 / 9.584.27 / 3.3026,184 / 23,7188,127 / 3,08434,311 / 26,802
EvoSkillGen90.7 / 86.24.63 / 3.081.42 / 1.473,184 / 4,5921,438 / 2,7814,622 / 7,373

GEdit, ImgEdit and RISEBench: 200 test tasks each. All columns use the same 600 tasks and report per-request means.

Thank you for the detailed response and the additional experiments. However, two concerns remain, and I will keep my original score.

  1. Limited contribution of key components. The new ablation in Table R7 shows that removing the Prompt Compiler lowers performance by only 0.8/0.5 points in the Strong T2I/Edit settings, suggesting that this component brings limited gains and is not clearly essential.

  2. Limited novelty. I agree with Reviewer g7dg that the paper's core ideas have already been explored in prior work and are not novel.

Reviewer_g7dgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents EvoSkillGen, which represents agentic workflows for image generation and editing as executable skills with applicability conditions. The system evolves a skill bank through skill selection, temporary skill construction, and cross-task skill revision. The authors integrate six existing benchmarks into GenAgentEval to evaluate the framework in terms of final-image quality, workflow-time overhead, and execution-trace diagnostics.

Critical strengths
  1. The paper represents agentic workflows for image generation and editing as executable skills with applicability conditions.
  2. The experimental scope is reasonable, covering six benchmarks, strong and weak agent configurations, and two judge models.
Critical weaknesses
  1. The paper's core ideas—composing workflows from basic operations, accumulating and revising a skill bank through execution feedback, and retaining only changes validated as effective—are not novel. These ideas have been explored in prior work. The main contribution therefore remains at the level of combining mechanisms and engineering implementation. The paper also lacks generalizable insights with practical explanatory value.
  2. The evidence for the performance advantage is not sufficiently convincing. The text-to-image lead is primarily driven by SpatialGenEval, whose evolution-stage checklists directly use the official spatial questions and target answers. This closely aligns the training feedback with the final scoring criteria. Moreover, apart from GenRouter, the other baselines are not given comparable access to training tasks and feedback-based adaptation from the same distribution. The comparison therefore does not adequately distinguish gains from the framework itself from gains due to training and adaptation to the target distribution.
  3. Most improvements are small: the text-to-image macro-average across three benchmarks exceeds the strongest baseline by only 0.9–1.8 points. Each benchmark contains only 200 test tasks, and the paper does not report repeated independent generation and skill-evolution runs, confidence intervals, or significance tests. It therefore remains difficult to rule out random variation as an explanation for these differences.
  4. The ablation studies and analytical experiments remain insufficient. Key controls needed to isolate the contribution of the framework design are missing, and the existing experiments do not adequately explain why the individual design choices are effective. In addition, there are no cross-benchmark transfer experiments, making it difficult to assess whether the learned skills generalize to benchmarks not encountered during evolution.
  5. The writing is insufficiently clear, and the exposition lacks logical coherence.
Questions for authors

See Weaknesses

5 comments
AuthorsOpenReview ↗

Author Response (1/3)

W1/W2. Contribution and controlled adaptation

Thank you for asking us to clarify novelty and explanatory findings. EvoSkillGen evolves a Skill bank from task execution experience so that learned Skills can be reused across tasks. Each Skill is an image-generation or editing workflow with explicit applicability conditions. Execution feedback guides changes to workflow structure and node instructions; image-quality comparisons select bank updates.

Original ablations show that Full exceeds independent per-request workflow construction by 3.0–6.5 points (R6); our new frozen-bank evaluation also finds gains on unseen GenEval and GIE-Bench (R9). The following controls examine where these benefits come from.

1. Structural changes add value beyond instruction optimization.

Our new R1 control starts both conditions from the same RS-Gen workflow and instructions: instruction-only evolution versus EvoSkillGen (structure + instruction evolution). Both share 300 evolution, 60 acceptance and 600 test tasks, original checklist sources, agents and generator. Model weights, meta-operation interfaces, Prompt Compiler and Router/Creator/Proposer role prompts stay fixed. Both retain Direct and share revision replay, pairwise acceptance, scope-generation and bank-management rules, and candidate-revision limits; banks are frozen before testing.

Instruction-only evolution allows multiple instruction variants, but all candidates from creation, revision, distillation and merging preserve RS-Gen's structure. EvoSkillGen also changes operations and dependencies. Thus, both receive feedback-based adaptation on the same distribution (Appendix F.7).

Table R1. Evolution from strong workflows and an instruction-only control.

Initial workflowMethod / settingWISEDPG-BenchSpatialGenEvalAvg.
RS-GenOriginal (Table 1)88.1 / 84.291.2 / 91.582.9 / 79.987.4 / 85.2
RS-GenInstruction-only evolution87.3 / 85.192.6 / 91.781.5 / 80.687.1 / 85.8
RS-GenEvoSkillGen (structure + instruction evolution)90.6 / 87.994.2 / 92.587.4 / 83.190.7 / 87.8
GEMSOriginal (Table 1)83.8 / 80.491.7 / 90.784.5 / 79.386.7 / 83.5
GEMSEvoSkillGen (structure + instruction evolution)88.2 / 86.293.9 / 92.386.6 / 83.889.6 / 87.4
DirectEvoSkillGen (structure + instruction evolution)87.6 / 84.392.8 / 91.487.2 / 82.689.2 / 86.1

Cells show Strong / Weak. Original-method and Direct-initialized scores are from Table 1. Image averages both judges; Avg. equally weights three benchmarks. RS-Gen and GEMS evolve separately. EvoSkillGen is our method.

EvoSkillGen scores 90.7/87.8 versus 87.1/85.8 for instruction-only evolution: 3.6/2.0 Strong/Weak points higher, leading on every benchmark under both agents. This supports structural changes under shared data, feedback, bank-update rules and candidate-revision limits.

The gains extend to non-spatial tasks: WISE improves by 3.3/2.8 and DPG-Bench by 1.6/0.8 Strong/Weak points. Averaging these two benchmarks alone gives 2.45/1.80 points from R1's displayed scores.

2. Skill evolution learns from a minimal start and improves strong workflows.

From Direct alone, EvoSkillGen raises the T2I average from 81.6 to 89.2/86.1 under Strong/Weak agents. From RS-Gen and GEMS, it improves over their original methods by 3.3/2.6 and 2.9/3.9 points, respectively; both evolved banks exceed Direct-initialized EvoSkillGen. Strong initial workflows and Skill evolution can therefore complement each other.

3. Workflow refinement and Skill-bank validation each contribute.

In original R6 ablations, admitting temporary Skills without Proposer editing loses 2.1–3.6 points. In new R7 controls, removing bank-level image comparison loses 3.8–6.1 points despite retaining revision replay and task-local RefineLoop. These results support workflow refinement and bank validation beyond Skill admission and local repair (Part 3).

Relation to prior work. GEMS uses trajectory memory to refine prompts and loads domain Skills; GenRouter updates memories for routing among predefined workflow-generator plans; AFlow searches code-represented workflows using execution feedback. EvoSkillGen evolves a bank of reusable Skills through structural and instruction changes selected by image-quality comparisons. R1 and R6–R9 provide controlled evidence for these choices and Skill reuse.

AuthorsOpenReview ↗

Author Response (2/3)

W2/W5. Supervision and execution boundaries

Thank you for highlighting supervision and reliability. Full is the main Direct-initialized EvoSkillGen.

Request-derived checklists (R2). We re-evolve Direct-only banks with checklists derived only from requests and required input images. Evolution feedback and acceptance both exclude official pre-generated fields: DPG questions and atomic requirements, SpatialGenEval questions/answers and explanations, and benchmark-specific editing criteria. Tasks, task order and other evolution settings remain unchanged; final banks are frozen and use the same test tasks and scoring protocol.

Table R2. Re-evolution with request-derived checklists.

SettingOriginal-checklist FullRequest-derived checklist
Strong T2I89.288.6
Weak T2I86.184.9
Strong Edit72.470.7
Weak Edit69.368.0
Strong SpatialGenEval87.286.7
Weak SpatialGenEval82.681.2

Full modality scores are from Table 3; SpatialGenEval averages both judges in Table 1.

Request-only SpatialGenEval scores are 86.7/81.2 (Strong/Weak), above the corresponding strongest baselines in Table 1: GEMS's 84.5 / RS-Gen's 79.9. Removing official fields lowers Full's spatial scores by 0.5/1.4 points, but its advantage persists. T2I averages are 88.6/84.9, versus original RS-Gen's 87.4/85.2: R2 supports spatial gains under both agents and the overall T2I advantage under Strong without official checklist fields.

Skill creation, updates and reuse. During evolution, L1 selects a Skill or creates a temporary one when a specialist is needed and none applies. Prompt Compiler integrates node artifacts with original inputs; RefineLoop runs when specified. After each batch, L2 revises, distills, evaluates and consolidates Skills. Acceptance and testing create no temporary Skills; unmatched requests use Direct. Testing freezes the bank.

W3. Reliability and practical value

We add independent workflow executions and Skill-bank evolution runs.

Independent workflow executions (R3–R4). Full and the corresponding strongest baseline in Tables 1–2 each run complete workflows three times per task through independent API requests: 600 test tasks per modality, 200 per benchmark. Full uses its final frozen banks. Strong/Weak use GPT-5.5/Doubao-Seed-2.0-pro with Seedream-4. Gemini 3.5 Flash and GPT-5.5 each repeat the complete scoring protocol three times per output.

R3 averages ratings and executions within each task, then bootstraps paired task differences stratified by benchmark, preserving equal benchmark, judge and ImgEdit-subset weights, to obtain 95% confidence intervals.

Table R3. Task-level paired statistical analysis.

SettingComparatorFull − baseline95% CI
Strong T2IRS-Gen+2.11[0.68, 3.62]
Weak T2IRS-Gen+1.03[−0.71, 2.73]
Strong EditMind-Brush+1.89[−0.36, 4.18]
Weak EditRS-Gen+3.33[0.69, 5.88]

R4 reports the mean and sample SD of three complete-run means, including agent execution, image generation and scoring.

Table R4. Independent complete-workflow executions with a fixed Skill bank.

SettingComparatorFull mean ± SDBaseline mean ± SD
Strong T2IRS-Gen89.03 ± 0.4086.92 ± 1.03
Weak T2IRS-Gen86.64 ± 0.7385.61 ± 0.53
Strong EditMind-Brush71.50 ± 1.7769.61 ± 1.90
Weak EditRS-Gen70.47 ± 2.1867.14 ± 2.36

Full retains higher mean scores in all four settings; editing has greater run-to-run variation than T2I.

Independent bank evolution (R5). We repeat evolution three times per modality from Direct-only banks: 300 evolution tasks in ten batches of 30, plus 60 acceptance tasks. Each final bank is frozen and evaluated once on the same 600 test tasks.

Table R5. Independent Skill-bank evolution.

SettingFull mean ± SD
Strong T2I89.87 ± 1.72
Weak T2I86.21 ± 1.06
Strong Edit71.32 ± 2.34
Weak Edit70.08 ± 3.11

Means differ from R4 by 0.18–0.84 points; SDs are higher in all four settings. R3 measures paired gains with final banks fixed; R5 measures Full's repeatability from evolution through execution and scoring.

Human preference and response time. R10–R11 use Strong outputs on 60 sampled tasks per modality, with three independent blinded judgments per pair. EvoSkillGen ranks first among three systems in both modalities, matching the automated ranking. Against Strong's corresponding strongest baselines (Tables 1–2), win/tie/loss counts are 97/31/52 vs RS-Gen (T2I) and 109/20/51 vs Mind-Brush (editing).

For Strong T2I, EvoSkillGen versus RS-Gen achieves 89.2 versus 87.4 Image points, 52.3 versus 120.5 seconds including image generation, 1.52 versus 2.68 image calls and 4,794 versus 11,156 total LLM tokens per request (R14; Appendix F.8).

AuthorsOpenReview ↗

Author Response (3/3)

W4. Design controls and cross-benchmark generalization

Thank you for requesting these controls. Original R6 ablations and new R7–R9 experiments examine workflow refinement, bank validation, simpler alternatives and Skill reuse.

1. Workflow refinement adds value beyond Skill admission

Table R6. Skill-learning ablations, score points.

VariantStrong T2IWeak T2IStrong EditWeak Edit
L1 only83.783.165.964.2
w/o Revision86.785.168.867.3
w/o Distillation85.684.069.167.0
w/o Consolidation87.785.471.068.2
Full89.286.172.469.3

R6 reports results from Table 3. Scores in R6–R8 equally weight three benchmarks per modality and both Gemini/GPT judges; ImgEdit weights Basic/UGE equally. Strong/Weak denote agents.

Full exceeds independent per-request workflow construction (L1 only) by 3.0–6.5 points. The w/o Distillation condition still admits temporary Skills but removes Proposer editing before admission, losing 2.1–3.6 points. Retaining workflows alone therefore does not recover the benefit of refining them from execution feedback. Removing revision or consolidation also lowers scores in all four settings.

2. Local repair and bank validation serve different roles

Acceptance controls retain revision replay and task-local RefineLoop. Compiler bypass fixes the bank and node outputs, appends text artifacts to the unchanged request in a fixed order, and uses the same image interface and generator settings.

Table R7. Prompt Compiler and selective-acceptance ablations.

VariantStrong T2IWeak T2IStrong EditWeak Edit
Full89.286.172.469.3
w/o Prompt Compiler88.484.371.967.0
L2 acceptance without bank-level image comparison84.182.366.365.2
Separate rubric scoring in L288.885.571.267.4

Full scores are from Table 3. Bank-level image comparison adds 3.8–6.1 points with local checks retained, supporting bank validation beyond local repair. Separate rubric scoring trails Full by 0.4–1.9 points, favoring pairwise acceptance. Compiler bypass costs 0.8/0.5 points for Strong T2I/editing and 1.8/2.3 for Weak.

3. Simpler inference workflows

R8 shares Full's agents, generator, test tasks and scoring protocol. Appendix F.1 includes rubric scoring and thinking-off rewriting.

Table R8. Intermediate baselines.

MethodStrong T2IWeak T2IStrong EditWeak Edit
Direct81.681.663.063.0
Single-call prompt rewriting (thinking off)83.181.864.363.1
Single-call prompt rewriting (thinking on)84.283.365.164.5
Direct + RefineLoop84.883.567.165.0
L1 only83.783.165.964.2
Full89.286.172.469.3

Direct is the baseline without agent assistance (Direct-Edit for editing), averaging Image scores in Tables 1–2; L1 only and Full are from Table 3.

Full exceeds thinking-enabled rewriting by 2.8–7.3 points and Direct + RefineLoop by 2.6–5.3 points. These results support learned workflows' image-quality benefits; R1 separately controls feedback-based adaptation.

For rewriting, GPT-5.5 uses reasoning_effort=none/medium; Doubao uses thinking.type=disabled/enabled, with reasoning_effort=medium when enabled. Both use max_completion_tokens=8192. Direct + RefineLoop follows Full's correction and stopping settings for up to five rounds. Candidates must win both anonymous presentation-order comparisons and, for editing, preserve required source content at least as well as the current image. Further details are in our response to 9rqo W3.

4. Unseen benchmarks

New R9 evaluates frozen Skill banks on complete official GenEval and GIE-Bench task sets, without evolution, acceptance or model selection on the target benchmarks. We use Seedream-4 at 1K with adaptive aspect ratios, official GenEval evaluation and GIE-Bench masks for non-edited-region CLIP similarity. GIE correctness uses GPT-5.5 in place of the official GPT-4o judge.

Table R9. Generalization to unseen benchmarks.

MethodGenEval (%) ↑GIE correctness (%) ↑GIE CLIP ↑
Direct89.5070.930.9107
EvoSkillGen (Strong)93.8377.780.9690
EvoSkillGen (Weak)92.3474.170.9512

Strong/Weak agents are GPT-5.5/Doubao-Seed-2.0-pro. Gains over Direct are 4.33/2.84 points on GenEval, 6.85/3.24 on GIE correctness and 0.0583/0.0405 in CLIP similarity. These gains without Skill-bank evolution on the target benchmarks support Skill reuse on unseen benchmarks, complementing R1 and R6–R7.

Reviewer_mGrYOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

EvoSkillGen represents image-generation workflows as reusable Skills consisting of operations, instructions, and artifact dependencies. For each task, the system either retrieves an existing Skill or creates a temporary one. It then uses execution feedback to revise existing Skills, turn successful temporary workflows into reusable Skills, and merge or remove redundant entries. The experiments cover three text-to-image and three editing benchmarks. The authors test both GPT-5.5 and Doubao-Seed-2.0-pro as agents while keeping Seedream-4 fixed as the generator. EvoSkillGen generally achieves the best average image-quality score and relatively low workflow overhead

Critical strengths

The method is well motivated. It is reasonable to represent reusable agent behavior as explicit executable workflows instead of unconstrained textual memory. The division between temporary task adaptation and longer-term Skill-bank evolution is also clean.

The evaluation is broader than usual for this type of agent paper. It includes six benchmarks, two agents with different capabilities, disjoint evolution and test sets, two judges, trace diagnostics, timing measurements, and several ablations. The full method performs consistently well across these settings.

Critical weaknesses

The main issue is that many improvements over the strongest baseline are only a few points, yet the paper reports no confidence intervals, paired tests, or variation across generation seeds. Running the judge three times on the same images measures judge variation, but not generation variance.

I am also unsure how much the conclusions depend on the chosen generator and evaluators. The main experiments use only Seedream-4. In addition, automated feedback is used both to evolve the Skill bank and to measure final performance. One final judge is GPT-5.5, which is also the strong agent. Human evaluation would help establish that the gains are not specific to this feedback ecosystem

Questions for authors
  1. Are the improvements over the strongest baseline statistically significant at the task level?

  2. What is the system ranking under blinded human evaluation?

  3. How expensive is the complete evolution process, including candidate replay and acceptance?

  4. Do Skills learned with Seedream-4 transfer to a different generator without additional evolution?

4 comments
AuthorsOpenReview ↗

Author Response (1/2)

Q1/W1. Statistical reliability and independent runs

Thank you for distinguishing judge variation from generation variance. We add independent complete-workflow executions and paired task-level analysis.

Full and the corresponding strongest baseline in Tables 1–2 each execute their complete workflows three times per task through independent API requests: 600 test tasks per modality, 200 per benchmark. Full uses its final frozen Skill banks. Strong/Weak use GPT-5.5/Doubao-Seed-2.0-pro with Seedream-4. Gemini 3.5 Flash and GPT-5.5 each repeat the complete scoring protocol three times per output. Scores equally weight both judges, benchmarks and ImgEdit subsets.

Task-level paired analysis. R3 uses unrounded Image scores, averaging repeated ratings and executions within each task. We bootstrap paired task differences stratified by benchmark, preserving ImgEdit-subset weights, to obtain 95% confidence intervals.

Table R3. Task-level paired statistical analysis.

SettingComparatorFull − baseline95% CI
Strong T2IRS-Gen+2.11[0.68, 3.62]
Weak T2IRS-Gen+1.03[−0.71, 2.73]
Strong EditMind-Brush+1.89[−0.36, 4.18]
Weak EditRS-Gen+3.33[0.69, 5.88]

Independent workflow runs. R4 reports the mean and sample SD of the three complete-run means, measuring variation from agent execution, image generation and scoring with the final banks fixed.

Table R4. Independent complete-workflow executions with a fixed Skill bank.

SettingComparatorFull mean ± SDBaseline mean ± SD
Strong T2IRS-Gen89.03 ± 0.4086.92 ± 1.03
Weak T2IRS-Gen86.64 ± 0.7385.61 ± 0.53
Strong EditMind-Brush71.50 ± 1.7769.61 ± 1.90
Weak EditRS-Gen70.47 ± 2.1867.14 ± 2.36

Full retains higher mean scores in all four settings. Editing shows greater run-to-run variation than T2I.

Q2/W2. Evaluator dependence and human evaluation

Because GPT-5.5 performs internal checks and update comparisons and also serves as a final judge, we add blinded human evaluation of the Strong (GPT-5.5) outputs.

Protocol. We sample 20 existing test tasks per benchmark across all six benchmarks (60 per modality; ImgEdit: 10 Basic and 10 UGE), using the original final outputs. Thirty non-author raters experienced in image processing provide three independent judgments per image pair. Raters see the original request, two outputs and, for editing, the source image. System identities, scores and feedback are hidden; left/right order is randomized. Choices are A / tie / B based on instruction satisfaction, including source preservation for editing.

We select the two strongest baselines per modality by equally averaging the two judges' Avg. Image scores in Tables 1–2: RS-Gen/GEMS for T2I and Mind-Brush/RS-Gen for editing. All three system pairs are compared, including the baseline–baseline pair.

Results. EvoSkillGen ranks first among the three systems in both modalities, matching the automated ranking on the full test set. Auto. scores average the two judges' displayed Avg. Image values. Preference scores assign 1/0.5/0 to wins/ties/losses, average the three judgments per task-pair, then equally weight three benchmarks and two opponents.

Table R10. Human and automated evaluation of T2I and editing.

ModalitySystemAuto. scoreAuto. rankHuman rankPreference score
T2IRS-Gen87.40220.5028
T2IGEMS86.65330.3417
T2IEvoSkillGen89.20110.6556
EditingMind-Brush70.10220.4597
EditingRS-Gen68.90330.3528
EditingEvoSkillGen72.40110.6875

Each R11 row covers 180 judgments: 60 tasks with three raters each.

Table R11. Pairwise vote counts for T2I and editing.

ModalityComparison (left vs right)Left winsTiesRight wins
T2IEvoSkillGen vs RS-Gen973152
T2IEvoSkillGen vs GEMS1132146
T2IRS-Gen vs GEMS1002753
EditingEvoSkillGen vs Mind-Brush1092051
EditingEvoSkillGen vs RS-Gen1113534
EditingMind-Brush vs RS-Gen883359
AuthorsOpenReview ↗

Author Response (2/2)

Q3. Complete evolution cost

Thank you for suggesting that we report the full cost of Skill-bank evolution and test Skill transfer without further evolution.

Table R12 reports one complete Skill-bank evolution run for each agent and modality, from the first task to the final bank update, including candidate replay and acceptance.

The accounting covers initial execution and feedback, temporary-Skill creation, candidate proposals, revision replay, bank acceptance, merging, pruning and retries, including failed and rejected candidates. Evolution time is the elapsed time from start to finish; API calls, LLM tokens and API fees are totals over this scope. Independent final test scoring is excluded from these totals.

Table R12. Cost of one complete Skill-bank evolution run.

SettingEvolution time (h)Other API callsGeneration/edit API callsLLM in/out tokensAPI cost (USD)
Strong T2I15.217,5952,11688.43M / 17.91M1,037.41
Weak T2I26.123,1342,575100.82M / 14.63M147.45
Strong Edit24.925,3262,80396.27M / 18.41M1,109.18
Weak Edit22.129,8283,283117.36M / 14.49M171.14

M denotes one million tokens. Other API calls cover LLM, search and tool services.

Table R12 shows that complete evolution takes 15.2–26.1 hours and 2,116–3,283 generation/editing API calls. API fees are USD 1,037.41 and 1,109.18 for Strong T2I and editing, respectively, versus USD 147.45 and 171.14 for Weak. Weak makes more calls in both API categories but incurs lower total API fees. Elapsed time varies by modality: Strong is faster for T2I, while Weak is faster for editing.

These are Skill-bank evolution costs; testing freezes the bank and reuses its Skills. The paper's Time column reports per-request workflow processing time, excluding image generation or editing.

Q4/W2. Transfer across generators

Table R13 compares Skill banks learned with Seedream-4 against Direct on Nano Banana 1 and Qwen-Image-2.1, without additional evolution. For this transfer experiment, neither target generator is used for Skill-bank evolution or model selection. We fix Skill contents, applicability conditions and Router/Prompt Compiler instructions, replacing the generator and adapting its interface. Testing creates no temporary Skills or bank updates. Unmatched requests use Direct, and RefineLoop retains its original settings.

Direct denotes the baseline without agent assistance; for editing, it corresponds to Direct-Edit in the main tables. It uses the original request and any required source image. We retain GPT-5.5 (Strong) and Doubao-Seed-2.0-pro (Weak) agents. Both methods use the same 600 test tasks per modality (200 per benchmark). Scoring follows the paper's cross-generator experiment: GPT-5.5 repeats each benchmark's complete protocol three times, with equal benchmark and ImgEdit-subset weights.

Table R13. Frozen Skill-bank transfer to two generators.

Target generatorMethodStrong T2IWeak T2IStrong EditWeak Edit
Nano Banana 1Direct86.286.265.265.2
Nano Banana 1Frozen-bank transfer89.787.667.566.0
Qwen-Image-2.1Direct88.188.173.473.4
Qwen-Image-2.1Frozen-bank transfer91.789.978.175.6

Nano Banana 1 Direct scores are from the paper's Table A3.

In Table R13, transfer improves over Direct by 3.5/1.4 points for text-to-image generation and 2.3/0.8 points for editing on Nano Banana 1 (Strong/Weak). The corresponding gains on Qwen-Image-2.1 are 3.6/1.8 and 4.7/2.2 points. The frozen Skill banks continue to outperform Direct on both target generators.

Dear Reviewer mGrY,

Thank you again for the time and effort you have devoted to reviewing our paper. As the discussion period draws to a close, we would like to ask whether our rebuttal, additional experiments, and manuscript revisions have adequately addressed your concerns. If any questions remain, we would be happy to provide further clarification. If our responses have resolved your concerns, we would be grateful if you could reconsider your score.

Best regards, The Authors

Thanks for the response from the authors. Considering the concerns from other reviewers, I will maintain my score.

Comments on the paper

1 comment

Dear Reviewers mGrY,

As the discussion phase is coming to a close, I encourage you to continue engaging with the authors until the discussion reaches a clear outcome. If any of your questions or concerns remain unaddressed, now is the time to raise them.

If there are no further questions, please respond to confirm your score after rebuttal.

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.