LinTree Enhances LLM Reasoning with Structured Search Histories
A new arXiv preprint (2605.31492) introduces LinTree, a method that improves large language model (LLM) reasoning by explicitly structuring search histories. The paper argues that LLMs generate intermediate traces akin to linearized search trees, where the model extends, abandons, and backtracks on partial solutions. Unlike traditional heuristic-guided search, this approach conditions on the entire search trace rather than just the local state. Experiments in Blocks World, grid Navigation, and Sokoban show that raw access to search history alone does not reliably outperform heuristic search. LinTree addresses this by structuring the history to better guide reasoning. The work highlights the potential of explicit search history structuring for enhancing LLM performance on complex reasoning tasks.
Key facts
- LinTree is introduced in arXiv:2605.31492.
- LLM reasoning traces are viewed as linearized search trees.
- The method conditions on the whole search trace.
- Tested in Blocks World, grid Navigation, and Sokoban.
- Raw search history access does not reliably beat heuristic search.
- LinTree structures search histories to improve reasoning.
- The paper is a new submission to arXiv.
- It focuses on improving LLM reasoning through explicit search history structuring.
Entities
Institutions
- arXiv