OpenClaw-Skill: Collective Skill Tree Search for Agentic Large Language Models

A team of researchers has proposed Collective Skill Tree Search (CSTS), a framework designed to automatically construct reusable skills for Large Language Model (LLM) agents, according to a paper submitted in 2026 [1]. The approach uses collective intelligence from multiple models to build structured skill trees aimed at improving performance on complex, real-world tasks [1]. The framework, detailed in a paper on arXiv, targets the challenge of equipping LLM agents with the capabilities needed for tool use, multi-step reasoning, and interaction within dynamic environments [1]. The resulting model, named OpenClaw-Skill, is built for the OpenClaw system [1]. The core mechanism of CSTS operates through two iterative phases: Collective Skill Node Generation (CSN-Gen) and Collective Skill Node Assessment (CSN-Assess) [1]. CSN-Gen taps the collective knowledge of multiple models to explore a diverse range of candidate skills for each subtask [1]. Following generation, CSN-Assess deploys multiple models as judges to evaluate and select the most promising skill nodes [1]. This assessment relies on two distinct scoring mechanisms [1]. The first, collective quality scoring, aggregates independent evaluations to produce a robust estimate of a skill's effectiveness [2]. The second, collective transferability scoring, explicitly verifies whether a skill can generalize well across different models [2]. Beyond the tree construction, the researchers introduce Collective Skill Reinforcement Learning [1]. This technique actively selects multiple relevant skills from the constructed tree to broaden the exploration of possible solutions [1]. The goal is to prevent the agent from becoming trapped by a single skill, which can lead to homogeneous or suboptimal outcomes [1]. The paper states that the fully trained OpenClaw-Skill model demonstrates strong agentic capabilities in long-horizon planning, tool use, and generalization when tested on challenging benchmarks [1].