SAR: A Structure-Aligned Reasoning Framework for Temporal Knowledge Graph Question Answering

Large language models (LLMs) augmented with retrieval have shown impressive performance in open-domain question answering, yet struggle significantly with temporal knowledge graph question answering (TKGQA). The core issue lies in structural misalignment: performing searches on structured, temporally sensitive knowledge graphs using plaintext queries often retrieves semantically similar yet structurally incorrect facts, resulting in critical inaccuracies. To address this, we introduce SAR, a Structure-Aligned Reasoning framework. SAR leverages an iterative agent-based architecture composed of three core modules: Reasoning and Answer Generation, Structure-Aligned Evidence Retrieval, and Iterative Answer Verification. The retrieval module is particularly essential; it employs structured query decomposition, embedding-based semantic matching, and chronological reranking to retrieve temporally consistent and schema-aligned knowledge facts from temporal knowledge graphs (TKGs). These precisely retrieved facts guide the LLM-based reasoning agent through iterative reasoning cycles, significantly reducing hallucinations and ensuring accuracy. A final verification stage ensures that proposed answers strictly adhere to requirements, reinforcing accuracy and coherence. Extensive experiments conducted on two benchmark datasets, MultiTQ and CronQuestions, demonstrate the effectiveness of SAR. Specifically, utilizing GPT-4.1, SAR achieves a Hits@1 score of 78.2% on MultiTQ, significantly surpassing existing methods, and similarly establishes new performance benchmarks on CronQuestions. Our findings highlight the crucial importance of structural alignment in temporal reasoning, particularly for complex queries involving multiple temporal constraints and multi-hop reasoning.