SPyCE: Skill-Policy Co-evolution for Multimodal Agents
Merged summary
TL;DR — SPyCE is a training framework for multimodal (image-reasoning) agents that distills tool-use trajectories into a reusable, hierarchical skill library that co-evolves with the RL policy, rather than collapsing experience into scalar rewards or relying on static test-time retrieval.
- Proposes "Skill-Policy Co-evolution": trajectories are distilled into reusable skills that are updated continuously throughout RL training, forming a closed loop where better policies yield better skills and vice versa.
- Uses a two-level skill hierarchy: execution skills for local visual operations and workflow skills encoding high-level priors that orchestrate multi-step tool use.
- During rollouts the policy conditions on retrieved skills for guidance, while the skill library is refreshed using high-value rollouts the policy generates.
- Reports consistent gains over both RL-based and memory-based baselines across eight benchmarks, with ablations indicating both the hierarchical skill design and co-evolution mechanism are necessary.
Sources (1)
SPyCE: Skill-Policy Co-evolution for Multimodal Agents
TL;DR — SPyCE is a training framework for multimodal (image-reasoning) agents that distills tool-use trajectories into a reusable, hierarchical skill library that co-evolves with the RL policy, rather than collapsing experience into scalar rewards or relying on static test-time retrieval.
- Proposes "Skill-Policy Co-evolution": trajectories are distilled into reusable skills that are updated continuously throughout RL training, forming a closed loop where better policies yield better skills and vice versa.
- Uses a two-level skill hierarchy: execution skills for local visual operations and workflow skills encoding high-level priors that orchestrate multi-step tool use.
- During rollouts the policy conditions on retrieved skills for guidance, while the skill library is refreshed using high-value rollouts the policy generates.
- Reports consistent gains over both RL-based and memory-based baselines across eight benchmarks, with ablations indicating both the hierarchical skill design and co-evolution mechanism are necessary.