Researchers have introduced a new framework to address the challenges of long-horizon reasoning in Large Language Models (LLMs), specifically focusing on "cross-skill" tasks that require seamless transitions between different reasoning abilities. To quantify this difficulty, the authors propose Skill Entropy, a metric that measures the complexity of switching between distinct skills within a reasoning chain. They developed Skill^2-Bench, a comprehensive benchmark comprising tasks across 9 verifiable and open-ended domains, utilizing a library of 558 distinct skills. Evaluation of 12 frontier and open-source models via Skill^2-Bench demonstrated a significant skill-switching gap, where performance accuracy declined as task-level skill entropy increased. To mitigate this, the authors developed Skill-Entropy RL, a reinforcement learning framework that rewards models for both step-level correctness and the alignment between their predicted skill sequences and the ground-truth skill sequences. Experimental results on Qwen3-4B-Instruct and Qwen3-1.7B showed substantial improvements, with Skill^2-Bench scores rising from 34.4% to 68.4% and 14.6% to 40.1%, respectively. The study demonstrates that skill entropy serves as a robust, reusable training signal that can be applied to existing datasets like OpenR1-Math to enhance multi-step reasoning capabilities.