TimeThink is a synthetic framework designed to elicit compositional reasoning in timeseries large language models (TS‑MLLMs). Existing TS‑MLLMs leverage LLM reasoning for QA tasks but often provide only implicit reasoning, failing to capture dynamic temporal patterns—an issue especially critical in high‑risk domains such as healthcare. Reinforcement‑learning‑based timeseries language models aim for explicit reasoning, yet they are limited by narrow in‑distribution training data and struggle with out‑of‑distribution compositional queries. TimeThink defines domain‑independent core timeseries primitives (e.g., trend, seasonality) that can be generated deterministically. Its synthetic data generator creates atomic and composite question‑answer pairs together with objective ground‑truth reasoning traces. Building on this, TimeThink employs a reinforcement learning with verifiable rewards (RLVR) strategy that incentivizes the model to produce explicit reasoning steps. Unlike template‑reliant methods, RLVR enables the model to learn the underlying logic of composition rather than merely mimicking traces. Extensive experiments show that a model trained solely on synthetic data with TimeThink significantly outperforms strong baselines on both synthetic and real‑world benchmarks.
Review: The study highlights the promise of combining synthetic data with verifiable reward signals, offering a new direction for interpretable reasoning in timeseries large language models.