Branch2Skill: Efficient Skill Evolution Through Reasoning Trees

Branch2Skill addresses the inefficiency of traditional skill evolution methods, which typically rely on single, linear reasoning trajectories. When a model makes an early error in a single rollout, subsequent feedback is often corrupted or sparse. Branch2Skill replaces this with a tree-based approach: for every training task, it performs Monte Carlo tree search (MCTS) to explore multiple reasoning paths simultaneously. By comparing an elite path with sibling nodes that share the same prefix, the system extracts localized, step-wise evidence about which reasoning patterns to retain, replace, or delete. This dense supervision is then distilled into a single, reusable update for the agent's persistent skill set.

Across six diverse benchmarks—including reasoning tasks like LiveMath and agentic tasks like ALFWorld—Branch2Skill consistently outperforms existing methods like SkillOpt. By consolidating evidence from multiple branches, the framework achieves higher task performance while drastically reducing the computational overhead. Specifically, when using GPT-5.5, Branch2Skill achieved a 9.3% increase in aggregate gain while consuming 73.2% fewer tokens compared to previous state-of-the-art methods. This demonstrates that reasoning trees provide a more efficient mechanism for extracting actionable lessons than repeated, isolated rollout-update cycles.

This research highlights a shift from treating reasoning as a linear sequence to viewing it as a structured search space. By leveraging the tree topology to isolate the impact of specific decisions, Branch2Skill enables agents to learn more effectively from their mistakes without needing exhaustive, costly retraining. This approach makes persistent skill evolution more practical for complex, multi-step tasks where early-stage errors often derail the entire reasoning process.