Dynamic path reasoning with large language models for interpretable knowledge graph question answering

Abstract Knowledge graph question answering (KGQA) aims to answer natural language questions by grounding them in structured knowledge graphs (KGs). Current KGQA systems still face three linked challenges: representing complex and compositional questions, providing transparent multi-hop reasoning traces, and using complete path-level evidence rather than only local triples. We propose SK-DRP ( S yntax- K nowledge enhanced D ynamic R easoning with P rompt learning), a modular KGQA framework that combines syntax-aware representation, state-adaptive path reasoning, and LLM-assisted answer selection. The syntax- and knowledge-enhanced encoder is an auxiliary representation enhancement that uses dependency information and knowledge-augmented pre-training to align questions with KG entities and relations. Here, dynamic denotes hop-wise recomputation of the question-focus indicator, score-qualified frontier expansion, and hop-weighted answer aggregation within a fixed dataset-level maximum hop budget; it does not denote automatic per-question depth selection. Candidate global paths are then compressed and transformed into natural-language auxiliary evidence for candidate-set-guided LLM answer selection. Experiments on MetaQA, WebQSP, and CWQ show that SK-DRP improves over several path-based and retrieval-based baselines on WebQSP and CWQ, while remaining competitive on MetaQA. Ablation, error, and interpretability analyses clarify how the syntax encoder, dynamic reasoning module, and LLM-assisted answer selection contribute to the final results, as well as the scope and limitations of the approach.