X-Tree
Preprint · 2026

X‑Tree: Tokenizing Reusable Experience
for Efficient Agent Generalization

Sitao Cheng1, Xunjian Yin2, Zhiyuan Sun1, Yuxuan Li1, Ruiwen Zhou3, Xiangru Jian1, Victor Zhong1

1University of Waterloo  ·  2Duke University  ·  3National University of Singapore

Contact: sitao.cheng@uwaterloo.ca · victor.zhong@uwaterloo.ca

Paper soon Code in progress Mined trees in progress Explore real trees

TL;DR X-Tree captures how a reusable skill is composed from sub-procedures. It is mined from agent trajectories with zero LLM calls, and it guides efficient training in three settings: as training instances for offline RL, as an adaptive bonus for online RLVR, and as privileged context of the self-teacher for OPSD.

Read the abstract

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, which build a vocabulary by counting alone, we score action spans by reusability and merge canonicalized actions into a reusable eXperience tree (X-Tree). Each X-Tree node captures how a frequent and success-bearing skill is composed from sub-skills, guiding efficient generalization.

We integrate X-Tree into three training settings: offline RL, with each node as a training instance; online RLVR, with an adaptive skill bonus; and on-policy self-distillation, with X-Tree as the self-teacher's privileged context. Across WebArena, ScienceWorld and WebShop at three model scales, X-Tree improves over standard recipes at matched data and budget by up to 4.5% SR on WebArena, 5.8% SR on ScienceWorld and 4.1% success on WebShop.

WebArena · offline RL+4.5SR over full-data SFT on the same trajectories
ScienceWorld · RLVR+4.9SR over outcome-only GRPO
ScienceWorld · OPSD+5.8SR over outcome-only RL, with X-Tree as the teacher's context
Building the tree0LLM calls: mined by deterministically measuring reusability
Interactive

Real trajectories, tokenized by the X‑Tree

Each example is a successful trajectory from a training corpus, with the tree mined from that corpus. Click a node to see the steps it covers and how often it recurs across the corpus. Drag the merge depth to replay the compression.

Merge depth

The tree drawing needs a wider screen. The bands, the merge stream and the node card show the same structure.

top-level skill next top-level skill primitive left unmerged node reused in this trajectory Node names are glosses written from each node's actions; the miner assigns only IDs.
Interactive

Where X‑Tree enters training

The same tree is used three ways. Step through each one to see when a node is triggered and what it changes.

Overview

Mine the hierarchy (X‑Tree) latent in the data and guide training with X‑Tree

Training with experience. (a) Prior training weights every token uniformly and harness systems keep LLM-written skills in context. (b) X-Tree mines reusable experience into a hierarchy with zero LLM calls and trains it into the weights as data, reward and context.
Training with experience. (a) Prior training weights every token uniformly, and harness systems keep LLM-written skills in context, never in the weights. (b) X-Tree mines reusable experience by X-Score into a hierarchy with zero LLM calls, and trains it into the weights as data (offline RL), as reward (online RLVR) and as context (OPSD).
Method

Tokenizing reusable experience

A text tokenizer merges co-occurring characters into a vocabulary. X-Tree applies the same idea to actions, with two changes: actions are first made comparable, and merges are chosen by how reusable the result is.

Step 1

Canonicalize

Each raw action becomes a typed token verb⟨role⟩. Element ids and object names are stripped into slots, so structurally identical actions share a symbol.

fill(bid, '2/2/26') → type⟨date⟩
Step 2

Score by X-Score

A pair of adjacent symbols is reusable when it recurs, when it is long, and when it appears in successful episodes. The pair with the highest score merges first.

recurrence × length × success
Step 3

Merge into a tree

A merge must also compress the corpus. Merging repeats until no pair passes or a cap is reached. Each node's depth is one more than its deeper child.

merge if fuv − 1 > η(ℓu + ℓv)
X-Score(u, v) = fuvrecurrence · (ℓu + ℓv)pℓlength · (succ(uv) + ε)pssuccess
Three training integrations with X-Tree: offline RL with nodes as RL instances, online RLVR with an adaptive node bonus, and OPSD with retrieved X-Tree skills as the self-teacher's privileged context.
Three training integrations. (a) Offline RL: each node is one RL instance, rolled out from the gold prefix and rewarded for step matching and node completion scaled by depth. (b) Online RLVR: a bonus for every node a rollout executes, weighted by λg, which is at full strength when no rollout in the group succeeds and fades as the verifier becomes informative. (c) OPSD: the same weights read the rendered X-Tree as privileged context for the self-teacher.
Results

Same data, same budget, more from each trajectory

Offline RL on WebArena: the tree drives the gain, not extra compute

With the same trajectories and no environment during training, offline RL on X-Tree nodes improves most on the procedure-heavy sites. Two more SFT epochs, random spans or an outcome-only reward recover little of it.

Per-site SR (%) on 694 WebArena tasks, Qwen2.5-7B. Go-Browse is our reproduction; every RL row starts from SFT on half the pool.
MethodgitlabshoppingadminredditmapwikiAll
Without RL
Qwen2.5-7B-Instruct (base)6.47.87.72.49.02.46.5
Go-Browse (SFT-Full, 2 epochs)16.821.722.916.410.47.118.4
  with 4 epochs (matched compute)17.519.424.416.113.411.918.8
SFT + offline RL
Offline RL, outcome reward15.523.423.115.27.514.318.2
Offline RL, random span18.420.926.614.711.914.319.5
SFT + offline RL + X-Tree
X-Tree, plain mixing16.124.525.419.914.423.820.7
X-Tree, depth curriculum17.821.528.620.812.423.821.2
X-Tree, full recipe22.026.629.318.810.414.322.9

Online RLVR: X‑Tree helps at every scale, including held-out tasks

The adaptive bonus improves outcome-only GRPO on seen and held-out ScienceWorld tasks and on WebShop success and graded score, at 1.5B, 3B and 7B. It helps most when the verifier signal is scarce.

Online RLVR bar chart: on ScienceWorld and WebShop, adding the X-Tree bonus to outcome-only GRPO improves every scale and fold.
ScienceWorld SR on three generalization levels and WebShop success and graded score. Each pair of bars changes only the reward. Means over three seeds.

OPSD: a mined tree teaches as well as an LLM-written skill bank

As the self-teacher's privileged context, the X-Tree rendering performs on par with skill banks written by gpt-oss-120b and GPT-o3, with no LLM calls. Its gain over outcome-only RL grows with model scale.

OPSD bar chart comparing outcome-only RLVR, the X-Tree bonus, OPSD with an LLM-written skill bank, and OPSD with the X-Tree bank.
Outcome-only RLVR, the X-Tree bonus, and OPSD with an LLM-written or an X-Tree skill bank. Only the skill bank the self-teacher reads changes. Means over three seeds.
Cite

BibTeX

@article{cheng2026xtree,
  title   = {X-Tree: Tokenizing Reusable Experience for
             Efficient Agent Generalization},
  author  = {Cheng, Sitao and Yin, Xunjian and Sun, Zhiyuan and
             Li, Yuxuan and Zhou, Ruiwen and Jian, Xiangru and
             Zhong, Victor},
  journal = {arXiv preprint},
  year    = {2026}
}