TimeThink
A synthetic data framework for temporal multimodal large language models. It deterministically generates atomic- and composite-level temporal question-answer pairs, and employs a reinforcement learning with verifiable rewards (RLVR) training strategy to enable compositional reasoning and improve generalization on out-of-distribution problems.
Timeline (last 90 days)
The key insight underlying TimeThink is that the core primitives of time-series data (e.g., trend, seasonality) are domain-agnostic and can be generated deterministically.
TimeThink designed a synthetic data generator that produces atomic-level and composite-level question-answer pairs, each accompanied not only by the correct answer but also by a complete reasoning trace.
TimeThink, trained solely on synthetic data, significantly outperformed strong baseline models on both synthetic benchmarks and real-world benchmarks.
TimeThink is a newly released arXiv preprint.
TimeThink bypasses the bottleneck of real-world annotation by deterministically generating verifiable synthetic data, offering a low-cost and scalable path.
All Facts (5)
The key insight underlying TimeThink is that the core primitives of time-series data (e.g., trend, seasonality) are domain-agnostic and can be generated deterministically.
50%UnverifiedTimeThink designed a synthetic data generator that produces atomic-level and composite-level question-answer pairs, each accompanied not only by the correct answer but also by a complete reasoning trace.
50%UnverifiedTimeThink, trained solely on synthetic data, significantly outperformed strong baseline models on both synthetic benchmarks and real-world benchmarks.
50%UnverifiedTimeThink bypasses the bottleneck of real-world annotation by deterministically generating verifiable synthetic data, offering a low-cost and scalable path.
50%UnverifiedTimeThink is a newly released arXiv preprint.
50%