Beyond Local Evidence: LLM-Guided Knowledge Graph Reasoning for Grounded Question Answering

ACL ARR 2026 May Submission14772 Authors

Keywords: Question Answering, Machine Learning for NLP, NLP and Symbolic Reasoning

Abstract: Large language models (LLMs) are increasingly combined with knowledge graphs (KGs) for grounded question answering. Existing LLM-KG methods mainly follow two paradigms: KG-RAG based generation uses KG evidence as auxiliary context, while LLM-guided KG reasoning constrains reasoning to traceable symbolic KG paths. However, current LLM-guided KG reasoning methods typically search within local neighborhoods of given topic entities, and may miss crucial evidence involving semantically relevant but non-adjacent entities, especially in incomplete KGs. To address this limitation, we propose LM-KGQA, an LLM-guided Monte Carlo Tree Search (MCTS) framework for expanding KG evidence search beyond local topic-entity neighborhoods. We further introduce self-rewarded node evaluation and aggregate high-reward traces for evidence-grounded answer derivation. Experiments on three KGQA benchmarks show that LM-KGQA outperforms nine SOTA baselines across three categories, with nearly 30\% average relative improvement on CWQ across backbone LLMs. Further analyses validate the self-reward mechanism and show robust performance across test-time reasoning budgets.

Paper Type: Long

Research Area: Question Answering

Research Area Keywords: Question Answering, Machine Learning for NLP, NLP and Symbolic Reasoning

Contribution Types: NLP engineering experiment

Languages Studied: English

EMNLP 2026 AI Reviewing Experiment: no

Submission Number: 14772