AI
X-Tree: Tokenizing Reusable Experience for Efficient Agent Generalization
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 25/09/2026, 20:00:00 UTC
Briefing created: 03/10/2026, 08:25:36 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/10/2026, 18:18:25 UTC.
Briefings
This article proposes X-Tree, a tokenization method for reusable experience in multi-step agents. By counting action spans rather than treating them uniformly, the method identifies hierarchical routines and merges them into a tree structure. This approach improves performance by up to 4.5% success rate on WebArena, 5.8% on ScienceWorld, and 4.1% on WebShop across three model scales. The system operates without LLM calls, using only text tokenizers and offline/online RL settings.
Key points
- X-Tree tokenizes action spans by counting frequency and success-bearingness to identify hierarchical routines.
- Training on X-Tree improves success rates by up to 4.5% on WebArena and 5.8% on ScienceWorld compared to standard recipes.
Context and limitations
Results are limited to three model scales and three datasets (WebArena, ScienceWorld, WebShop). Performance gains are measured as success rate improvements, not absolute success rates. The study does not report generalization across different agent architectures or training environments beyond the tested settings.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 03/10/2026, 08:25:36 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeidnkoqtczqh4zkqxdlgo4krxadibvqgelu5ml2tt4flhinhm2rzfy