Many of the signals that matter most in the real world come as numerical time series: heart-rate and ECG traces in hospitals, stock prices and economic indicators in finance, or temperature and pollution measurements in environmental monitoring. Today, making sense of these streams usually means building narrow prediction models and then asking human experts to interpret what those numbers actually mean, why they changed, and what might happen next. At the same time, large language models (LLMs) like ChatGPT are very good at working with text: they can answer questions, explain their reasoning, and connect to background knowledge. However, they struggle to understand and reason about time series signals directly due to the mismatch between text and dense numerical data. This project aims to close that gap by developing AI systems that can natively understand, analyze, and predict numerical time series while talking about them in everyday language.
Our approach combines new ways of representing time series with training strategies that teach LLMs to reason step-by-step about temporal data. In one line of work (TsLLM [1]), we build effective and efficient time series perception into a language model by adding a specialized module that turns raw signals into a form the model can understand, and then training it on millions of paired examples where it must forecast, answer questions, describe patterns, classify events, and generate short reports about what it sees. In a second line of work (COUNTS [2]), we use reinforcement learning to encourage the model to learn reasoning strategies on its own that help it solve tasks involving time series data. The model learns to compute metrics, check patterns, explore “what-if” scenarios, and explain its conclusions – rather than jumping straight to an answer. Together, these contributions move beyond black-box prediction: they create models that can look at complex temporal data, explain what is happening and why, and support decision-makers in domains where both numerical accuracy and clear, human-readable explanations are critical.
