LLM-Based Time Series Question Answering with Review and Correction

This paper addresses the challenge of applying Large Language Models (LLMs) to time series question answering (TSQA). It highlights the limitations of existing LLM approaches in handling numerical sequences and proposes a novel framework, T3LLM, that leverages the inherent verifiability of time series data. The framework uses a worker, reviewer, and student LLMs to generate, review, and learn from corrected reasoning chains, respectively. This approach is significant because it introduces a self-correction mechanism tailored for time series data, potentially improving the accuracy and reliability of LLM-based TSQA systems.

• •Proposes T3LLM, a novel framework for time series question answering.
• •T3LLM utilizes a worker, reviewer, and student LLM architecture.
• •The framework incorporates a self-correction mechanism based on the verifiability of time series data.
• •Demonstrates state-of-the-art performance on TSQA benchmarks.