The strong performance of LLMs (Large Language Models) across various domains, our research indicates that factual questions involving temporal reasoning remains a significant challenge for these models. This challenge primarily arises from LLMs’ insensitivity to implicit time parameters, hindering their ability to accurately interpret the internal meaning and leads to erroneous answers. To address this, we propose the Auto Time-CoT framework. This framework integrates function calling, CoT (Chain-of-Thought) and rule-based methods to implement various data operations, including deconstruction strategy of multi-dimensional data, question reconstruction, automatic explicit of implicit time parameters and structured time information extraction. Additionally, we refine the implicit time parameters and adopt a divide-and-conquer approach, proposing various Manual-CoT methods. Auto Time-CoT leverages the inherent capabilities of LLMs without relying on external knowledge, enhancing their potential to answer factual questions involving temporal reasoning. We evaluate Auto Time-CoT on three datasets and find that LLMs have considerable potential for data processing, planning and coordinating various function calling of tools, achieving up to an 8% improvement.

Question-answering tasks are one of the most important tasks in natural language processing (NLP). Recently, large language models (LLMs) have demonstrated outstanding performance, but our experiments reveal that such factual questions involving time reasoning still pose a challenge for existing LLMs. This is primarily attributed to the insensitivity of LLMs to numbers and inability to comprehend the implicit meaning of time, leading to incorrect responses. In order to overcome these limitations, we propose Time Chain-of-Thought (Time-CoT). Specifically, Time-CoT represents the thought process in answering factual questions involving time reasoning and serves as a prompt for LLMs, including explicitation of the implicit "time parameter", the description of the "standard question entity" along the dimension of the timeline and reasoning based on the "time parameter". The aim is to stimulate its potential to answer such questions without external knowledge. We evaluate Time-CoT on three related question-answering datasets and achieve significant improvements, with the most significant improvement of up to 10.87% over strong baselines.