Reproduce “MemoTime: Memory-Augmented Temporal Knowledge Graph Enhanced Large Language Model Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xingyu Tan, Xiaoyang Wang, Qing Liu, Xiwei Xu, Xin Yuan, Liming Zhu, Wenjie Zhang
MemoTime introduces a memory-augmented framework that enhances large language models' temporal reasoning by integrating temporal knowledge graphs, hierarchical decomposition, and continuous experience learning. The system achieved state-of-the-art Hits@1 of 77.9% on the MultiTQ dataset and 71.4% overall accuracy on TimeQuestions, notably boosting smaller LLMs' performance to levels comparable with larger, untuned counterparts.
Key insights
- Grounding LLM reasoning in Temporal Knowledge Graphs (TKGs) with strict temporal faithfulness is essential for overcoming limitations in dynamic, time-evolving information processing.
- Hierarchical decomposition of complex temporal questions, facilitated by a "Tree of Time," effectively enforces monotonic timestamp progression and synchronizes multi-entity temporal contexts.
- A self-evolving experience memory enables continuous learning from successful reasoning traces, significantly improving efficiency, stability, and the ability to generalize across diverse temporal operators.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.