Chain-of-Timeline: Enhancing LLM Zero-Shot Temporal Reasoning with SQL-Style Timeline Formalization

Jiaying Wu, Bryan Hooi

Keywords: temporal reasoning, time-sensitive QA, Large Language Models

TL;DR: We introduce Chain-of-Timeline (CoTime), a training-free framework that enhances LLM temporal reasoning by distilling and structuring temporal facts into a formalized, SQL-style timeline.

Abstract: Accurate reasoning about time-sensitive facts is essential in today's rapidly evolving knowledge landscape. While Large Language Models (LLMs) possess impressive reasoning capabilities, they struggle with time-sensitive question answering (QA) in long documents due to the presence of (1) irrelevant noisy context and (2) implicit expressions of temporal events. To address these challenges, we introduce Chain-of-Timeline (CoTime), a framework that constructs topic-relevant event timelines through structured code-style formalization. CoTime first extracts a high-level topic from the question (e.g., [subject]'s career history) to identify relevant temporal events in the document. These events are then organized into a temporal SQL-style schema, enabling CoTime to derive answers based on the question's specified time identifiers. Experimental results show that CoTime surpasses representative baselines.

Submission Number: 121