.png){kind=icon}
Return
Temporal Knowledge Graph Forecasting via Constrained Large Language Models
.png){kind=icon}
PRE
AI
DOI：10.1109/tbdata.2026.3689005.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Temporal Knowledge Graph Forecasting (TKGF) remains challenging due to the complex temporal dependencies and relational interactions in evolving knowledge graphs. Recent studies have explored Large Language Models (LLMs) for TKGF through prompt engineering or iterative multi-step querying. While these approaches demonstrate promising reasoning capabilities, they often rely on unconstrained decoding and do not explicitly expose structured, temporally valid reasoning trajectories. To address these limitations, we propose T-CoLLM, a two-stage reasoning framework that explicitly decouples path generation from answer prediction. In the first stage, we fine-tune an LLM to generate candidate multi-hop reasoning paths under a dual-strategy guidance mechanism, where Trie-based constraints ensure temporal consistency and structural validity. In the second stage, a reasoning LLM predicts the final answer based on these candidate paths, followed by a structure-aware reranking module that reconciles semantic confidence with graph topology. Extensive experiments on multiple benchmark datasets demonstrate consistently competitive or superior performance compared to strong baselines. Further analyzes verify the effectiveness of each component and highlight the transparency provided by our explicitly constructed and temporally verifiable reasoning paths.
Keywords:
Temporal knowledge graphs
large language models
temporal knowledge graphs forecasting
Journal
I
IF:
5.7
Papers:
894
Citations:
3.0K
Organization
View more organizations
Cited Papers
.png){kind=icon}
ShareSaveShareSave
ShareSave View more