X-Tree: Tokenizing Reusable Experience for Efficient Agent Generalization

• 类型：arxiv
• 标识：2609.32993
• 链接：https://arxiv.org/abs/2609.32993
• 主分类：agent
• 形态：method
• TLDR：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. Recent agents do use that structure, but only as LLM-written skills in context, never in the weights, so their gains do not generalize beyond retrieval. We instead recover this hierarchy from the data itself and train on it, with no LLM calls. Following text tokenizers,
• 待LLM分类：否
• 标题中文：X-Tree：面向高效 Agent 泛化的可复用经验 token 化方法
• TLDR中文：多步 Agent 在扁平的 action 流上训练：SFT 和 RLVR 对每个 token 一视同仁地加权，忽略了跨任务反复出现的子流程以及使人类能够自顶向下基于可复用 routine 进行规划的层级结构。这种结构被闲置，而扁平训练对每条稀缺 trajectory 的利用程度低于其内容所允许的水平。近期 Agent 确实使用了这种结构，但仅作为由 LLM 编写、置于 context 中的 skill，从未体现在 weights 中，因此其收益无法泛化到检索之外。我们转而从数据本身恢复这种层级结构并据此训练，无需任何 LLM 调用。借鉴 text tokenizer 的思路……
• 来源文件：
• /inbox/tom/_candidates/2026-10-03-agent-rag-longcontext-candidates.json