acceptodds
Desk rejected at ICLR 2027

V-Gym: Enhancing Agentic Visual Reasoning via Skill-Data Co-Evolution

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Advances in multimodal understanding, reasoning, and tool use enable agents to tackle increasingly complex visual reasoning tasks. By distilling past execution experience into reusable skills, agents can transfer lessons from both successes and failures into future reasoning, reducing repeated errors and improving capabilities. However, limited experience may produce unreliable, poorly generalizable skills, while static datasets may lack the targeted and diverse practice needed for refinement. To address this gap, we introduce V-Gym, an autonomous framework that iteratively co-evolves procedural skills and multimodal practice data from execution trajectories. During skill evolution, V-Gym analyzes trajectories to distill and refine hierarchical skills, updating procedural guidance and applicability conditions while retaining an update only if it improves validation performance. During data evolution, V-Gym selects generation seeds by balancing data utility and exploration, then translates trajectory-identified bottlenecks into diverse, targeted practice data that expand the data bank after quality checks. The resulting practice outcomes feed back into subsequent skill updates, closing the loop for continual skill refinement. Experiments across diverse multimodal reasoning benchmarks show substantial improvements over baselines with multiple backbone models. Its evolved skills generalize across domains and models, while evolved data support more effective skill refinement, enabling autonomous diagnosis, targeted practice, and continual self-improvement.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Agent, Self-Evolution

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_qFxUOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  1. Method. The paper proposes V-Gym, a framework that improves a fixed-parameter multimodal agent for tool-assisted visual reasoning by co-evolving a hierarchical skill bank and a practice data bank from execution trajectories.
  2. Two components. (1) Skill evolution reflects on grouped trajectories to propose skill patches, which are accepted only if they improve performance on a validation set. (2) Data evolution attributes skill-update gains to practice instances, selects seeds with UCB-style scoring, and generates targeted new instances that preserve core task challenges.
  3. Results. Experiments on TIRBench, MMSearch-Plus, and MMBrowseComp show gains over existing skill-evolution methods, and the evolved skills transfer to unseen domains and weaker models.
Critical strengths
  1. Novel skill-data co-evolution. Using the same execution trajectories to both refine skills and generate targeted practice data is a well-motivated idea, and the internal comparisons show that targeted generation is more effective than generic synthesis under the same budget.
  2. Transferable skills. The evolved skills improve performance on unseen benchmarks and on weaker models without further updates, suggesting that the learned procedural knowledge is reusable.
Critical weaknesses
  1. Insufficient ablations. (1) The utility-based UCB selection is only compared with random selection. A simple heuristic, such as prioritizing failed or low-accuracy instances, would be a stronger control for justifying the utility attribution. (2) The self-check in data generation is not ablated; it remains unclear how much it contributes compared with admitting generated data directly. (3) The validation gate uses a large in-distribution labeled validation set (e.g., 630 instances vs. 195 training instances on TIRBench), whereas the w/o Validation Gate variant practices only on the training set. Adding the validation data to practice without the gate would help separate the effect of the gating mechanism from that of the additional labeled data.

  2. Limited baseline comparison. The skill-evolution baselines (XSkill, Ace-Skill, and SkillOPT) are evaluated only with GPT-5.5. With Gemini-3.5-Flash and Qwen-3.7-Flash, V-Gym is compared only with Vanilla Tools, which is insufficient to support the claim that it consistently outperforms baselines across backbones.

  3. Missing cost analysis. Only the training rollout budget (i.e., during skill evolution) is matched across methods, while the additional costs of V-Gym, including validation reruns, reflection, and data generation with self-checks, are not reported. (1) Time cost: the time cost of both training and test-time inference should be reported and compared with the baselines. (2) Computational cost: the number of API calls and token consumption in both phases are not reported.

Questions for authors

See weakness 1-3.

Reviewer_Qe8nOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes V-Gym, a framework that jointly evolves reusable skills and training data from agent execution trajectories to improve agentic visual reasoning, achieving better performance than existing baselines across several benchmarks and models.

Critical strengths
  1. Good performance
Critical weaknesses
  1. Limited novelty - Skill evolution, feedback-driven data generation, and co-evolution have already been explored in many prior works. Beyond the works discussed in the paper, several related papers have already been accepted, e.g., Meta Context Engineering (ICML 2026), EvoSkill (COLM 2026), CoEvolve (ACL 2026), LoopTool (ACL 2026), etc., but are not adequately discussed or compared against. The proposed method largely combines existing techniques, including trajectory reflection, validation-based skill updates, UCB sampling, and targeted data generation. Even the learned skills in Appendix D look like generic prompting instructions.
  2. Unfair experimental setup - figure 4 compares V-Gym against a baseline that repeatedly trains on the same real data and a generic synthetic data baseline with random seed selection and a fixed data mixture. A proper offline vs. online data generation comparison with matched data sizes, sampling strategies, and compute budgets is missing. The paper only matches training rollout budgets, without controlling for the total compute spent on data generation, quality checking, and skill validation. The ablations also fail to establish whether the learned skills provide meaningful benefits beyond standard prompting or task-specific planning. Overall, it is unclear whether the gains come from co-evolution, better data selection, more diverse data, or simply additional compute.
Questions for authors
  1. Can you clarify the novelty over existing works, focusing on new research insights and methodological contributions rather than engineering or implementation differences?
  2. Can you address the fairness concerns in your experimental setup, particularly the lack of controlled comparisons with matched data generation strategies and total compute budgets?
  3. How does V-Gym perform in OOD settings/datasets? e.g. other datasets like HR-MMSearch (CVPR 2026), etc.?
Reviewer_9KAbOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how visual agents can improve through practice with fixed model parameters. To study this, this paper proposed V-Gym to learn reusable skills from execution traces, and uses these traces to generate targeted practice data. Skill updates are preserved only when they improve validation results. Data selection considers both past utility and exploration. The experiments are conducted on three benchmarks and the results show gains over tool-only agents and existing methods. The learned skills can also transfer to other benchmarks and weaker models, which further validates the effectiveness.

Critical strengths
  1. The proposed method connects skill learning with targeted practice in a useful way.
  2. V-Gym outperforms the compared skill methods on all three benchmarks with GPT-5.5.
  3. The learned skills can transfer to new benchmarks and weaker models.
Critical weaknesses
  1. As shown in lines 322–323, the methods use the same training rollout budget. V-Gym also needs extra work to generate data and check skill updates. Whether it still performs better when the total cost is the same?

  2. In lines 223–236, the skill updates are tested one by one and then merged. I am not sure whether changes that work well alone still work well together. Does the method check the result after merging?

  3. In lines 399–400, the ablation uses random selection for both generation seeds and next-epoch practice data. This makes it hard to analyze which part contributes more to the gain. Therefore, testing them separately would make the result clearer.

Questions for authors

Please refer to the weakness.

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

Related papers

Discussion (0)

Sign in to comment.