Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TempR1: Improving Temporal Understanding of MLLMs via Temporal-Aware Multi-Task Reinforcement Learning

Tao Wu, Li Yang, Gen Zhan, Yabin Zhang, Yiting Liao, Junlin Li, Deliang Fu, Li Zhang, Limin Wang

TL;DR

TempR1 tackles the challenge of robust temporal understanding in multimodal large language models by training a single model across five interconnected temporal tasks with a unified, task-aware reinforcement learning objective. It introduces GRPO-based multi-task RL with a structured reward scheme, including a machine-parsable format reward and three-category temporal localization rewards to align predictions with ground truth. Leveraging a 60K multi-task corpus, TempR1 demonstrates state-of-the-art performance across TG, VHD, TAL, DTG, and GVQA benchmarks and shows clear synergistic gains from joint task optimization. This work provides a scalable, principled framework for temporal reasoning in MLLMs with broad generalization across diverse temporal scenarios.

Abstract

Enhancing the temporal understanding of Multimodal Large Language Models (MLLMs) is essential for advancing long-form video analysis, enabling tasks such as temporal localization, action detection, and time-sensitive question answering. While reinforcement learning (RL) has recently been explored for improving temporal reasoning, existing approaches are often confined to limited task types and data, restricting their generalization across diverse temporal understanding scenarios. To address this challenge, we present TempR1, a temporal-aware multi-task reinforcement learning framework that systematically strengthens MLLMs' temporal comprehension. We curate a multi-task corpus that exposes the model to diverse temporal structures and semantics, and build upon the Group Relative Policy Optimization (GRPO) algorithm to achieve stable and effective cross-task optimization. Specifically, we categorize temporal tasks into three correspondence types between predicted intervals and ground-truth instances, and design tailored localization rewards for each, enabling TempR1 to capture fine-grained temporal dependencies and adapt to different temporal patterns. Extensive experiments demonstrate that TempR1 attains state-of-the-art performance across multiple benchmarks. Moreover, its joint optimization over complementary tasks yields a strong synergistic effect, enhancing both generalization and single-task performance, establishing a scalable and principled paradigm for temporal reasoning in MLLMs.

TempR1: Improving Temporal Understanding of MLLMs via Temporal-Aware Multi-Task Reinforcement Learning

TL;DR

TempR1 tackles the challenge of robust temporal understanding in multimodal large language models by training a single model across five interconnected temporal tasks with a unified, task-aware reinforcement learning objective. It introduces GRPO-based multi-task RL with a structured reward scheme, including a machine-parsable format reward and three-category temporal localization rewards to align predictions with ground truth. Leveraging a 60K multi-task corpus, TempR1 demonstrates state-of-the-art performance across TG, VHD, TAL, DTG, and GVQA benchmarks and shows clear synergistic gains from joint task optimization. This work provides a scalable, principled framework for temporal reasoning in MLLMs with broad generalization across diverse temporal scenarios.

Abstract

Enhancing the temporal understanding of Multimodal Large Language Models (MLLMs) is essential for advancing long-form video analysis, enabling tasks such as temporal localization, action detection, and time-sensitive question answering. While reinforcement learning (RL) has recently been explored for improving temporal reasoning, existing approaches are often confined to limited task types and data, restricting their generalization across diverse temporal understanding scenarios. To address this challenge, we present TempR1, a temporal-aware multi-task reinforcement learning framework that systematically strengthens MLLMs' temporal comprehension. We curate a multi-task corpus that exposes the model to diverse temporal structures and semantics, and build upon the Group Relative Policy Optimization (GRPO) algorithm to achieve stable and effective cross-task optimization. Specifically, we categorize temporal tasks into three correspondence types between predicted intervals and ground-truth instances, and design tailored localization rewards for each, enabling TempR1 to capture fine-grained temporal dependencies and adapt to different temporal patterns. Extensive experiments demonstrate that TempR1 attains state-of-the-art performance across multiple benchmarks. Moreover, its joint optimization over complementary tasks yields a strong synergistic effect, enhancing both generalization and single-task performance, establishing a scalable and principled paradigm for temporal reasoning in MLLMs.

Paper Structure

This paper contains 15 sections, 11 equations, 4 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Revisiting Group Relative Policy Optimization
  5. Task Formulation and Format Reward
  6. Temporal Localization Reward
  7. Multi-task Training with GRPO
  8. Experiments
  9. Training Data
  10. Benchmarks and Evaluation Metrics
  11. Implementation Details
  12. Comparison with State-of-the-arts
  13. Ablation Study
  14. Qualitative Results
  15. Conclusion

Figures (4)

  • Figure 1: Performance comparison across five temporal understanding tasks. Our proposed TempR1 achieves the best overall results.
  • Figure 2: Overview of the TempR1 framework. We finetune the MLLM on a multi-task training corpus covering five temporal understanding tasks. Reinforcement learning is applied with rule-based rewards, including format and accuracy rewards, as well as localization rewards for three correspondence types: Type 1 (one-to-one, TG/DTG), Type 2 (many-to-one, VHD/GVQA), and Type 3 (many-to-many, TAL). These rewards jointly optimize the MLLM to achieve accurate and robust temporal prediction.
  • Figure 3: Comparison with the Qwen2.5-VL-7B base model and supervised fine-tuning (SFT) results on general video benchmarks.
  • Figure 4: Qualitative result comparisons. (a) Comparison of two matching strategies for localization reward in the TAL task, showing that the DP-based method yields more accurate interval alignment. (b) Comparison of TempR1 and VideoChat-R1 on NExTGQA, where TempR1 provides more precise visual evidence localization.