Skill is Not One-Size-Fits-All: Model-Aware Skill Alignment for LLM Agents

In three linesMASA (Model-Aware Skill Alignment) adapts procedural skills for LLM agents to each model backbone without weight modification. A hierarchical evolution pipeline rewrites skills via hill climbing and UCB-driven tree search, then a lightweight rewriter trained on trajectories reproduces adaptation in a single forward pass. Gains up to 25.8 points across three interactive environments and four backbones.

Summary generated by Claude — human-verified