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

Skill effectiveness in LLM agents varies significantly by model backbone — a skill that improves one model can degrade another. MASA addresses this with a two-stage framework: hierarchical skill rewriting followed by model-aware selection, adapting retrieved procedural instructions per backbone without modifying model weights. Tested across multiple model scales on long-horizon interactive tasks.