MCTS-KBQA: Monte Carlo Tree Search with Information Gain Rewards for Knowledge Base Question Answering

August 19, 2026 | Source: arXiv AI

Tags: knowledge base QA, Monte Carlo Tree Search, KBQA, LLM reasoning, information retrieval, arXiv

Fast MCTS replaces expensive rollout-based search in knowledge base question answering with a training-free information gain reward computed by a single LLM forward pass — consistently improving the accuracy-cost tradeoff on four KBQA benchmarks. Accepted at CIKM 2026.

Details

Monte Carlo Tree Search applied to LLM-based knowledge base question answering (KBQA) is computationally expensive because rollout-based search requires generating complete answer paths. Existing MCTS-style methods either rely on direct LLM scoring (unreliable) or separately trained reward models (data-intensive), and typically provide rewards only at terminal states. Fast MCTS replaces terminal rollouts with an information gain (IG) reward for intermediate states: a question-conditioned perplexity-ratio proxy computed over sanitized interaction histories using a single forward pass of an open-source instruction LLM. No additional reward model training is required. On four KBQA benchmarks, Fast MCTS consistently outperforms linear baselines and generally improves the accuracy-cost tradeoff compared to rollout-based Classic MCTS. The key advantage is that IG rewards are cheap to compute — making tree search tractable for complex multi-hop reasoning over knowledge graphs. Accepted to CIKM 2026. The approach's dependence on PPL computation from a specific open-source model may limit direct portability to API-only settings. Code and data are publicly available.