X-Tree: Tokenizing Reusable Experience for Efficient Agent Generalization
The paper proposes learning reusable agent “skills” from action trajectories and training models on that hierarchy directly.
X-Tree finds repeated, success-bearing action spans in agent data, canonicalizes them, and merges them into a tree of reusable experience without LLM calls. The authors plug that structure into offline RL, online RLVR, and on-policy self-distillation. In tests on WebArena, ScienceWorld, and WebShop, they report gains over standard training recipes at matched data and budget, topping out at 5.8% success-rate improvement on ScienceWorld. HF Daily Papers' note
X-Tree finds repeated, success-bearing action spans in agent data, canonicalizes them, and merges them into a tree of reusable experience without LLM calls. The authors plug that structure into offline RL, online RLVR, and on-policy self-distillation. In tests on WebArena, ScienceWorld, and WebShop, they report gains over standard training recipes at matched data and budget, topping out at 5.8% success-rate improvement on ScienceWorld. HF Daily Papers' note
score 5