Researchers have developed a new framework called NeST that improves the performance of large language models (LLMs) in time series forecasting. The framework, which stands for neighborhood-aware semantic alignment and temporal modulation, adapts LLMs to forecast continuous signals by integrating textual and time series information in the prompt. This is achieved through three steps: generating text prototypes optimized to represent local neighborhoods of pretrained word token embeddings, aligning them with temporal representations of the input using a nearest-neighbor contrastive objective, and introducing text prototype conditioned temporal modulation that adaptsively scales and shifts time series features. The authors claim that NeST outperforms state-of-the-art methods in eight benchmark datasets and demonstrates strong generalizability on real-world tasks. --- Why it matters: This matters to researchers in AI because it provides a new approach to adapting LLMs for time series forecasting, which is a challenging task due to the difference between discrete text data and continuous signals. The framework's ability to integrate textual and time series information could lead to improved performance on various real-world tasks. Source: https://arxiv.org/abs/2412.04806