Large language models are increasingly used for reasoning over time series data, but their typical approach is slow. They generate long chains of thought, which can consume thousands of tokens for a single analysis.
The new paper, posted on arXiv, proposes TimesJev, a model designed to "think fast" on time series. It uses a principle the authors call System 1 Sufficiency, drawing on the idea of fast, intuitive processing rather than slow, deliberate step-by-step reasoning.
By reducing reliance on extended reasoning chains, TimesJev aims to make time series reasoning more efficient. The paper's abstract does not include experimental details, but the motivation is clear: cut the computational cost while retaining the reasoning capability of LLMs.