PathFinder: MCTS and LLM Feedback-based Path Selection for Multi-Hop Question Answering

2025-12-04

AI Summary

In multi-hop question answering, large language models (LLMs) frequently suffer from hallucination and erroneous reasoning paths, hindering performance. To address this, we propose PathFinder—a novel framework that synergistically integrates Monte Carlo Tree Search (MCTS) with LLM-based feedback mechanisms. PathFinder employs MCTS to generate diverse reasoning trajectories; leverages an LLM-as-a-judge to evaluate and filter high-confidence paths using sub-answer recall; dynamically reconstructs failed sub-queries to enhance path robustness; and incorporates retrieval-augmented generation (RAG) for closed-loop optimization. Evaluated on multiple mainstream multi-hop QA benchmarks, PathFinder achieves significant accuracy improvements over strong baselines. It effectively mitigates hallucination, reduces erroneous reasoning, and minimizes retrieval failures. Our results empirically validate the efficacy of a search-and-feedback-coordinated, interpretable reasoning paradigm for complex multi-step QA.

Technology Category

• Search and Optimization: Metareasoning and Metaheuristics
• Machine Learning: Large Multimodal Models (LMMs)
• Multiagent Systems: Multiagent Learning

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Multi-hop question answering is a challenging task in which language models must reason over multiple steps to reach the correct answer. With the help of Large Language Models and their reasoning capabilities, existing systems are able to think and decompose an input question over multiple steps to analyze, retrieve, and reason. However, training-based approaches for this problem still suffer from LLM hallucinations and incorrect reasoning paths that hinder performance. Hence, we propose PATHFINDER, an approach that: (i) uses Monte Carlo Tree Search to generate training path traces, (ii) improves training data quality by filtering erroneous and lengthy traces using sub-answer recall and LLM-as-a-judge verification, and (iii) reformulates sub-queries to handle failed retrieval cases. By following these steps, we demonstrate that PATHFINDER improves the performance of multi-hop QA over public benchmark datasets.

Problem

• Improves multi-hop QA by filtering erroneous reasoning paths
• Uses MCTS to generate and refine training path traces
• Enhances retrieval handling through sub-query reformulation

Innovation

• Uses Monte Carlo Tree Search for path generation
• Filters training data with recall and LLM verification
• Reformulates sub-queries to handle retrieval failures