ABSTRACT Temporal knowledge graph reasoning (TKGR) aims to predict future facts based on events that have occurred up to the current timestamp in a temporal knowledge graph (TKG). Traditional learning‐based methods focus on structural information but overlook semantics inherent in TKGs, while rule‐based methods struggle to generate sufficient high‐quality temporal logical rules. Recently, large language models (LLMs) have been introduced into the TKGR task, typically leveraging retrieved relevant historical events and LLMs' semantic understanding and reasoning abilities to infer missing information in the query. Existing LLM‐based methods encompass both fine‐tuning and non‐fine‐tuning variants. The former demands high computational cost and exhibits limited transferability, whereas the latter suffers from performance constraints stemming from the absence of specialised adaptation training. In this paper, we introduce GT‐LLM, a novel non‐fine‐tuning TKGR framework that enhances performance by improving history retrieval quality while avoiding costly task‐specific LLM fine‐tuning and preserving flexible adaptability to different or dynamically updated TKGs. The key to our method is that it transforms a series of TKG subgraphs into a single, unified topological graph, where entities, relations, and timestamps are explicitly modelled as nodes and connected through topological edges to expose richer associations for history retrieval. Based on this unified representation, we design a general TKGR pipeline that performs history retrieval based on topological association strength and temporal relevance, followed by LLM reasoning under vocabulary‐level constraints. Furthermore, we integrate the predictions from the LLM with those from a graph‐based model to harness their complementary strengths. Extensive experimental results on standard TKGR benchmarks demonstrate that our method significantly improves history retrieval quality and achieves promising overall performance.
Chen et al. (Sun,) studied this question.