Fast MCTS: Information Gain Rewards for More Efficient LLM-based KBQA
Fast MCTS, introduced in arXiv paper 2502.13428v2, improves LLM-based knowledge base question answering by applying Monte Carlo Tree Search with information gain rewards. It replaces costly terminal rollouts with an intermediate-state reward signal, computed as a question-conditioned perplexity-ratio proxy over sanitized interaction histories. This requires only forward passes of an open-source instruction LLM, eliminating the need for separate reward-model training. The method consistently outperforms linear baselines across four KBQA benchmarks, addressing existing challenges of reward design and search expense.
Key facts
- The paper is arXiv:2502.13428v2 with announcement type replace-cross.
- It applies Monte Carlo Tree Search (MCTS) to LLM-based KBQA.
- Reward design is difficult and rollout-based search is computationally expensive in existing methods.
- Existing MCTS-style methods rely on direct LLM scoring or require substantial data to train separate reward models.
- Fast MCTS uses an information gain (IG) reward for intermediate states instead of terminal rollouts.
- The IG reward is a question-conditioned PPL-ratio proxy over sanitized interaction histories.
- It uses an open-source instruction LLM without additional reward-model training.
- Fast MCTS outperforms linear baselines on four KBQA benchmarks.
Entities
Institutions
- arXiv