LLMs Can Infer Events from Time Series Data

A recent study reveals that Large Language Models (LLMs) are capable of deducing natural language events from time series data. The researchers created an automated approach to generate tasks using sports data, evaluating 18 different LLMs. The findings indicate unexpected effectiveness, even when provided with limited context. Furthermore, the integration of distillation and Reinforcement Learning (RL) enhanced the performance of smaller models, enabling them to nearly match the capabilities of larger proprietary reasoning models. All resources utilized in the study are accessible to the public.