TimeMaster: Training Time-Series Multimodal LLMs to Reason via Reinforcement Learning

TL;DR

TimeMaster introduces a reinforcement learning approach for multimodal large language models to perform structured, interpretable reasoning over visualized time-series data, achieving state-of-the-art results on real-world classification tasks.

Contribution

It presents a novel RL-based training method with a structured output format for time-series reasoning in multimodal LLMs, improving performance and interpretability.

Findings

Outperforms classical models and GPT-4o on TimerBed benchmark.

Achieves over 14.6% and 7.3% performance gains.

Generates explanations and domain insights.

Abstract

Time-series reasoning remains a significant challenge in multimodal large language models (MLLMs) due to the dynamic temporal patterns, ambiguous semantics, and lack of temporal priors. In this work, we introduce TimeMaster, a reinforcement learning (RL)-based method that enables time-series MLLMs to perform structured, interpretable reasoning directly over visualized time-series inputs and task prompts. TimeMaster adopts a three-part structured output format, reasoning, classification, and domain-specific extension, and is optimized via a composite reward function that aligns format adherence, prediction accuracy, and open-ended insight quality. The model is trained using a two-stage pipeline: we first apply supervised fine-tuning (SFT) to establish a good initialization, followed by Group Relative Policy Optimization (GRPO) at the token level to enable stable and targeted reward-driven improvement in time-series reasoning. We evaluate TimeMaster on the TimerBed benchmark across six real-world classification tasks based on Qwen2.5-VL-3B-Instruct. TimeMaster achieves state-of-the-art performance, outperforming both classical time-series models and few-shot GPT-4o by over 14.6% and 7.3% performance gain, respectively. Notably, TimeMaster goes beyond time-series classification: it also exhibits expert-like reasoning behavior, generates context-aware explanations, and delivers domain-aligned insights. Our results highlight that reward-driven RL can be a scalable and promising path toward integrating temporal understanding into time-series MLLMs.