Long-horizon reasoning tasks often require models to switch between distinct skills, such as performing a math derivation and then using the result for planning. The authors of a new paper call these cross-skill long-horizon tasks and argue that existing benchmarks fail to measure how well models handle such transitions.
To fill this gap, they introduce Skill Entropy, a metric that quantifies the difficulty of switching between skills. They also build Skill^2-Bench, a benchmark with 558 skills across 9 verifiable and open-ended domains, where each task is assigned a skill-entropy score and grouped into three difficulty levels.
Evaluations on 8 frontier and 4 open-source models reveal a skill-switching gap: accuracy drops as task entropy increases. This suggests that current models struggle with high-entropy, cross-skill reasoning.
The authors then turn skill entropy into a training signal with Skill-Entropy RL, a reinforcement learning framework where the model predicts both the answer and the skill used at each step. The reward combines step-level correctness with a skill-entropy reward that aligns the predicted skill sequence with the gold sequence.
On Qwen3-4B-Instruct and Qwen3-1.7B, Skill-Entropy RL improves Skill^2-Bench scores from 34.4% to 68.4% and from 14.6% to 40.1%, respectively, outperforming competitive baselines. The approach also works on off-the-shelf data like OpenR1-Math, indicating its reusability. Code is available on GitHub.