LiTS: A Modular Framework for LLM Tree Search

LiTS is a new modular Python framework that enables LLM reasoning through tree search algorithms like MCTS and BFS. The framework demonstrates reusable components across different domains and reveals that LLM policy diversity, not reward quality, is the key bottleneck for effective tree search in infinite action spaces.

• →LiTS decomposes tree search into three reusable components: Policy, Transition, and RewardModel that work across different algorithms.
• →The framework was tested on MATH500, Crosswords, and MapEval tasks, showing components are reusable across algorithms within task types.
• →A key finding reveals that in infinite action spaces, LLM policy diversity is more critical than reward quality for effective tree search.
• →The framework is open-source under Apache 2.0 license with installation instructions and runnable examples.
• →Domain experts can extend to new domains through decorator-based registry system while researchers can implement custom search algorithms.