Thanks to the powerful performance of reasoning capabilities of DeepSeek-R1, reinforcement learning-based fine-tuning paradigms have garnered widespread attention from researchers. Some studies have explored the preliminary performance of GRPO in multimodal tasks, such as object localization, counting, etc. We investigate the potential of GRPO in the video temporal grounding task, which demands precise temporal alignment between visual and linguistic modalities as well as advanced reasoning capabilities. This task is particularly well-suited for our approach due to its reliance on fine-grained temporal dynamics, which facilitate the development of intuitive rule-based reward mechanisms and enable the model to iteratively refine its reasoning and outputs.
[2025/3/21] 🔥 The code and checkpoints has been released! Please check our huggingface repo. [Checkpoints]
- Training-Framework: We utilize the Easy-R1 framework and contribute to video training.
- Model: We select Qwen2.5-VL-3B as base model.
- Dataset: Charades and ActivityNet-tvg.
git clone https://github.com/appletea233/Temporal-R1.git
cd Temporal-R1
pip install -e .
# eval with lmms-eval
cd third_party/lmms-eval
pip install -e .
-
Download the annotation files and videos
-
You need to create a file named
tvg.yamlunderexamples/data_configwith the following content:
datasets:
- json_path: xxx.json
data_folder: xx
- json_path: yyy.json
data_folder: yy
The json_path is the dataset file, and the data_folder stores the videos.
Train the Model:
bash examples/qwen2_5_vl_3b_tvg.sh
Run Inference:
# Custom Inference
bash third_party/lmms-eval/examples/eval_tvg.sh $GPUS $MODEL_PATH $TASKS
# R1 Inference
bash third_party/lmms-eval/examples/eval_tvg_r1.sh $GPUS $MODEL_PATH $TASKS
# task uses temporal_grounding_charades,temporal_grounding_activitynet
| Para. | token | RL | think | mIoU(Charades) | mIoU(ANet-tvg, OOD*) | Checkpoint |
|---|---|---|---|---|---|---|
| 3b | 2048 | ❌ | ❌ | 37.22 | 18.92 | Qwen/Qwen2.5-VL-3B-Instruct |
| 3b | 2048 | SFT | ❌ | 45.95 | 20.86 | SFT-3B-Charades |
| 3b | 2048 | ✅ | ❌ | 51.10 | 22.10 | Temporal-R1-3B-Charades |
| 3b | 2048 | ✅ | ✅ | 53.93 (+7.98) | 23.07 (+2.21) | Temporal-R1-3B-Charades |
*OOD: Our model is trained exclusively on Charades-tvg, while ANet-tvg represents out-of-domain data.
Experimental results demonstrate that, compared to the SFT model, the GRPO-trained model not only achieves significant performance improvements but also exhibits reasoning ("think") capabilities and stronger generalization. Specifically, the mIoU on the Charades dataset increased by +7.98, while the mIoU on the ActivityNet benchmark also showed a improvement (+2.21). These findings indicate that GRPO training enables the model to perform better in complex tasks and adapt more effectively to diverse data distributions. In addition, we also evaluated the model's performance when generating only the final output without including its reasoning process. The experimental results indicate that performance declines across the board, suggesting that the inclusion of a reasoning process has a positive effect on our model. We plan to release more related experimental results in the future.
We also explored the performance of our model when directly tested on the VideoQA task using MVBench. Our model achieves an accuracy of 59.6, slightly lower than the base model's 63.35. However, the model fine-tuned through direct supervised fine-tuning on the same training data completely lost its ability to output valid options. This phenomenon highlights that reinforcement learning-based fine-tuning preserves a significantly higher degree of generalization compared to SFT.
-
Scale up model, datasets
-
Widen more downstream tasks, e.g., VideoQA, Temporal Referring, Video Captioning, etc.
We want to thank EasyR1, Qwen2.5-VL, llama-factory and lmms-eval for publicly releasing their code and pretrained models.
Contributors: Hongyu Li, Songhao Han, Yue Liao, Jialin Gao, Si Liu
@misc{li2025temporalr1,
title = {Envolving Temporal Reasoning Capability into LMMs via Temporal Consistent Reward},
author = {Hongyu Li, Songhao Han, Yue Liao, Jialin Gao, Si Liu},
howpublished = {\url{https://github.com/appletea233/Temporal-R1}},
year = {2025}
}