Daily Paper Briefing

Abstract Overview

ViSkill is a reinforcement learning framework for vision-language model agents that stores successful interactions as composite visual skill cards rather than reducing them to text. These visual-native skills preserve spatial layouts, action-state progression, and a distilled strategy description in a form directly consumable by the agent. During training and inference, the system retrieves skills using geometric similarity, uses them as visual context, and also derives an auxiliary skill-guided reward from them. Successful trajectories are then distilled back into the library, creating a closed feedback loop in which policy improvement and skill accumulation co-evolve, with an optional cold-start library to improve early exploration.

Novelty

The paper's main distinction is its visual-native representation of reusable skills as composite image-based skill cards, instead of text-centric skill memories that may lose geometric structure. It also tightly couples skill retrieval, reward shaping, and online skill distillation within reinforcement learning, so the skill library and policy improve each other during training.

Results

Across Sokoban, FrozenLake, and PrimitiveSkill, ViSkill reaches an overall success rate of 0.89, increasing to 0.91 with cold-start initialization, and exceeds all evaluated proprietary and open-source baselines in the reported overall metric. It reports 0.82 on Sokoban, 0.84 on FrozenLake, and 1.00 average success on PrimitiveSkill, while also converging faster than standard PPO. Ablations show that visual skill representations outperform text-based alternatives and that removing joint optimization causes the largest degradation.

Key Points

1. ViSkill represents successful experience as composite visual skill cards that preserve spatial and procedural information for VLM agents.
2. The framework links geometry-aware skill retrieval, skill-guided rewards, and online distillation into a closed feedback loop between the policy and the skill library.
3. Empirically, the method outperforms reported baselines overall and benefits further from cold-start initialization, with ablations supporting the value of visual skills and co-optimization.

References

• arXiv: https://arxiv.org/abs/2610.12403v1
• Fugu-MT: https://fugumt.com/fugumt/paper_check/2610.12403v1
• Hugging Face Papers: https://huggingface.co/papers/2610.12403
• GitHub: https://github.com/ZJU-REAL/ViSkill