X-Tree: Tokenizing Reusable Experience for Efficient Agent Generalization
X-Tree tokenizes reusable agent skills and raises success on WebArena, ScienceWorld, and WebShop.
X-Tree recovers hierarchical reusable skills from agent trajectories by scoring action spans for reusability and merging them into an experience tree, without LLM calls. Each node represents a frequent, success-bearing skill composed of sub-skills. The method is integrated into offline RL, online RLVR with an adaptive skill bonus, and on-policy self-distillation. Across three model scales it improves over matched baselines by up to 4.5% success rate on WebArena, 5.8% on ScienceWorld, and 4.1% on WebShop.
- X-Tree builds a reusable skill hierarchy from action spans without LLM calls.
- Each node captures a frequent, success-bearing skill composed of sub-skills.
- Integrated into offline RL, online RLVR, and on-policy self-distillation.
- Gains up to 4.5% SR on WebArena, 5.8% on ScienceWorld, and 4.1% on WebShop.
From the original article
1 min read at huggingface.co
Multi-step agents are trained on flat action streams: SFT and RLVR weight every token uniformly and ignore the sub-procedures that recur across tasks, the hierarchy that lets humans plan top-down from reusable routines. This structure sits unused, and flat training uses each scarce trajectory less fully than its content allows.
By Sitao Cheng, Xunjian Yin, Zhiyuan Sun · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.