Reproduce “TsLLM: Augmenting LLMs for General Time Series Understanding and Prediction” on OpenResearch

About this paper

Felix Parker, Nimeesha Chan, Chi Zhang, Kimia Ghobadi

Researchers at Johns Hopkins University developed TSLLM, a multimodal generative model that augments large language models with specialized time series perception, enabling high-fidelity, interleaved text-time series reasoning and generation. The model unifies diverse time series analysis tasks, such as forecasting and classification, through autoregressive generation, achieving competitive performance on traditional benchmarks while significantly advancing contextual question-answering and contextual forecasting.

Key insights

• Augmenting a pretrained LLM with dedicated time series perception and generation abilities allows for effective, integrated reasoning across numerical temporal signals and natural language.
• A β-VAE objective for time series encoding provides a beneficial semantic abstraction to the LLM, with regularized representations proving more effective for downstream tasks than near-perfect signal reconstruction fidelity.
• A multi-stage training strategy and extensive multimodal data curation, including synthetic generation, are crucial for developing robust, general-purpose time series understanding and generation capabilities in an LLM.

Summary by alphaXiv.