From Recognition to Understanding: Unlocking Cognitive Time Series Reasoning with LLMs

The article introduces TSCognition, a multimodal benchmark designed for multi-dimensional time series reasoning, comprising approximately 41K QA samples across five cognitive reasoning tasks. It also presents TSAlign, a framework that encodes time series into compact representations and aligns them with LLM embeddings using gated residual injection and multivariate fusion. Experimental results indicate that TSAlign significantly outperforms existing baselines on TSCognition and TimerBed while reducing computational costs, highlighting its potential for enhancing semantic understanding in time series analysis with LLMs.