Toward Skill-Native LLMs: Skill Entropy for Benchmarking and Training Long-Horizon Reasoning

This paper introduces Skill Entropy, a measure of how difficult it is for an LLM to switch between reasoning skills during long-horizon tasks. It presents Skill²-Bench, covering 558 skills across nine verifiable and open-ended domains, and evaluates eight frontier and four open-source models. Accuracy declines as task-level skill entropy rises. The proposed Skill-Entropy RL trains models to predict both intermediate answers and the skills used, reportedly raising scores for Qwen3-4B-Instruct from 34.4% to 68.4% and Qwen3-1.7B from 14.6% to 40.1%.