Temporal-R1 is a research project on reinforcement learning tuning (RLT) for video multimodal large language models (VideoLLMs), aiming to enhance the model's reasoning capabilities for video during the post-training stage. The project points out that understanding real-world videos with complex semantics and long temporal dependencies remains a fundamental challenge in computer vision; although multimodal large language models have demonstrated strong capabilities in vision-language tasks, reinforcement learning tuning further improves their reasoning abilities.
Main capabilities and design: Based on the Group Relative Policy Optimization (GRPO) framework, the project proposes a dual-reward formula that simultaneously supervises semantic reasoning and temporal reasoning through discrete and continuous reward signals. To support effective preference optimization, the project introduces a variance-aware data selection strategy based on repeated reasoning to identify samples that can provide effective learning signals. The project evaluates this method on eight representative video understanding tasks, including VideoQA, Temporal Video Grounding, and Grounded VideoQA.
Experimental results: The official description states that this method consistently outperforms supervised fine-tuning (SFT) and existing RLT baselines, achieving better performance with significantly less training data. In the Video Temporal Grounding results, compared with the SFT model, the GRPO-trained model not only shows significantly improved performance, but also demonstrates reasoning ("think") ability and stronger generalization ability; specifically, mIoU on the Charades dataset improves by +7.98, and mIoU on the ActivityNet benchmark also improves (+2.21). In addition, when evaluation generates only the final output without including the reasoning process, performance declines overall, indicating that including the reasoning process has a positive effect on the model. When directly testing on MVBench for the VideoQA task, the model accuracy is 59.6, slightly lower than the base model's 63.35; however, a model directly supervised fine-tuned on the same training data completely lost the ability to output valid options. This phenomenon indicates that reinforcement learning-based fine-tuning retains significantly higher generalization than SFT.
Usage: The project provides installation guidelines, requiring downloading annotation files and videos, and creating a file named tvg.yaml under examples/data_config, in which json_path is the dataset file, and data_folder stores videos. The command for training the model is given in the repository.
Release status: The initial version of the code was released on March 21, 2025, and points to the Hugging Face repository (Checkpoints); the paper (Paper) was released on June 2, 2025. The official notice indicates that a new version of the code including an optimized reward mechanism and additional datasets will be released soon.
Applicable scenarios: Research and experimental scenarios for video understanding and reasoning tasks such as video temporal grounding, video question answering, and Grounding VideoQA.
Limitations and unavailable information: The project does not provide information such as fees, pricing, commercial licensing, platform compatibility scope, or complete supported languages; details such as the specific training command, the content of the tvg.yaml example, and the dataset source and scale should be based on the repository files and are not expanded in this introduction.
