CSTS framework builds reusable skill trees to boost LLM agent performance
Researchers Tianyi Lin, Chuanyu Sun, and Jingyi Zhang propose Collective Skill Tree Search (CSTS), a tree-search framework that automatically constructs structured, reusable skill trees to improve LLM agents in tool use, multi-step reasoning, and dynamic environment interaction.
Score breakdown
CSTS addresses a core bottleneck in agentic LLM development by replacing manual skill engineering with an automated, multi-model collective process that explicitly tests whether skills transfer across models — a property the paper identifies as critical for robust generalization.
- 01The paper proposes Collective Skill Tree Search (CSTS), a tree-search-based skill construction framework for LLM agents.
- 02CSTS operates through two iterative phases: Collective Skill Node Generation (CSN-Gen) and Collective Skill Node Assessment (CSN-Assess).
- 03CSN-Gen uses collective knowledge from multiple models to explore diverse candidate skills for each subtask.
Tianyi Lin, Chuanyu Sun, and Jingyi Zhang present Collective Skill Tree Search (CSTS), a framework designed to automatically construct reusable skills for LLM agents operating in complex real-world systems like OpenClaw. The core mechanism runs two iterative phases: CSN-Gen, which draws on collective knowledge from multiple models to explore diverse candidate skills for each subtask, and CSN-Assess, which uses multiple models as judges to score skill nodes on two dimensions — collective quality scoring (aggregating independent evaluations for a robust estimate of skill effectiveness) and collective transferability scoring (explicitly verifying whether a skill generalizes across different models). The output is a comprehensive tree of skills paired with skill-augmented training data that enables models to learn and apply those skills effectively.
Building on the skill tree, the paper introduces Collective Skill Reinforcement Learning, a training approach that actively selects multiple relevant skills from the tree during learning.
Building on the skill tree, the paper introduces Collective Skill Reinforcement Learning, a training approach that actively selects multiple relevant skills from the tree during learning. This multi-skill selection is designed to broaden solution-space exploration and prevent the model from converging on a single skill's homogeneous or suboptimal solutions. The trained model, OpenClaw-Skill, is reported to exhibit outstanding agentic capabilities across long-horizon planning, tool use, and generalization over challenging benchmarks.
Key facts
- 01The paper proposes Collective Skill Tree Search (CSTS), a tree-search-based skill construction framework for LLM agents.
- 02CSTS operates through two iterative phases: Collective Skill Node Generation (CSN-Gen) and Collective Skill Node Assessment (CSN-Assess).
- 03CSN-Gen uses collective knowledge from multiple models to explore diverse candidate skills for each subtask.
- 04CSN-Assess scores skill nodes on collective quality (aggregated independent evaluations) and collective transferability (cross-model generalization).
- 05The framework also introduces Collective Skill Reinforcement Learning, which selects multiple relevant skills to broaden solution-space exploration.
- 06The resulting trained model is called OpenClaw-Skill, targeting long-horizon planning, tool use, and generalization.
- 07Authors are Tianyi Lin, Chuanyu Sun, and Jingyi Zhang, published on ArXiv (2606.16774).
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 16, 2026 · 23:11 UTC. How this works →