TimeThink Elicits Compositional Reasoning in Timeseries Language Models

Summary

Timeseries multimodal large language models can answer questions but often fail to represent dynamic temporal patterns explicitly, which limits the explanations needed in high-stakes settings. Existing reinforcement-learning approaches are also prone to narrow training distributions and poor performance on out-of-distribution compositional questions. TimeThink addresses this problem with a synthetic framework built around domain-independent primitives such as trends and seasonality. It generates atomic and composite question-answer pairs with deterministic ground truth and reasoning traces. The framework then applies reinforcement learning with verifiable rewards to encourage explicit reasoning and to learn the logic of composition rather than imitate templates. Experiments reported in the paper show that a model trained only on synthetic data outperforms strong baselines on both synthetic and real-world benchmarks.