Research map / Video / Video-language understanding
Video-R1: Reinforcing Video Reasoning in MLLMs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Video-R1 applies R1-style rule-based RL with a new temporal-contrast GRPO variant (T-GRPO) and a mixed image-video reasoning dataset to improve video reasoning in Qwen2.5-VL-7B.
Problem
Directly applying GRPO-style RL to video reasoning in MLLMs lacks explicit temporal modeling, so models shortcut via single frames, and high-quality video reasoning training data is scarce.
Approach
T-GRPO adds a temporal contrastive reward: correct responses get a bonus when accuracy on temporally ordered frames is at least accuracy on shuffled frames. Training also mixes image reasoning data with video data, with an SFT cold start on CoT-annotated data.
Key results
- Video-R1-7B reaches 37.1% on VSI-Bench at 64 frames, above GPT-4o's 34.0%.
- Video-R1-7B at 64 frames scores 52.4 on VideoMMMU and 64.8 on MVBench, versus 50.4 and 59.2 for Qwen2.5-VL-7B (CoT) at 64 frames.
- Ablations show removing image data (Video-R1-7B-wo-image) drops VSI-Bench from 34.6 to 32.3 at 16 frames; the T-GRPO-trained model incorporates temporal reasoning in 75.0% of responses versus 60.2% without it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add image-based training data (ablation) | Video-R1-7B · VSI-Bench | accuracy: 32.3 -> 34.6 | Helped |
| add image-based training data (ablation) | Video-R1-7B · VideoMMMU | accuracy: 45.8 -> 49.8 | Helped |
| use T-GRPO instead of GRPO (ablation) | Video-R1-7B · VSI-Bench | accuracy: 32.7 -> 34.6 | Helped |
| use T-GRPO instead of GRPO (ablation) | Video-R1-7B · VideoMMMU | accuracy: 48.3 -> 49.8 | Helped |
| include an SFT cold start before RL (ablation) | Video-R1-7B · VSI-Bench | accuracy: 31.8 -> 34.6 | Helped |
| include an SFT cold start before RL (ablation) | Video-R1-7B · VideoMMMU | accuracy: 49.5 -> 49.8 | No clear effect |
| increase input frames from 16 to 64 | Video-R1-7B · VSI-Bench | accuracy: 34.6 -> 37.1 | Helped |
| increase RL training from 1k to 10k steps | Video-R1-7B (32 frames) · MMVU (mc) | accuracy: 63.8 -> 66.2 | Helped |
| add temporal reward (T-GRPO) (ablation) | Video-R1 · temporally grounded responses across six benchmarks | percentage of responses with temporal reasoning: 60.2% -> 75.0% | Helped |
| add length reward (ablation) | Video-R1 · six benchmarks | average performance | Helped |
| vary temporal reward magnitude α (ablation) | Video-R1 · six benchmarks | average performance | Mixed |
Surprising findings
- SFT on the CoT data did not consistently improve over the base model; on VideoMME performance slightly dropped after SFT at some frame counts.
Limitations
- Training uses only 16 frames, which limits long-range temporal modeling.
- T-GRPO adds computational overhead from the contrastive ordered and shuffled rollouts.
- Length reward uses a fixed interval regardless of question difficulty.
Experimental setup
Datasets: Video-R1-CoT-165k Video-R1-260k
Benchmarks: VSI-Bench VideoMMMU MMVU (mc) MVBench TempCompass VideoMME (wo sub)
Metrics: accuracy
Baselines: GPT-4o LLaVA-OneVision-7B Video-UTR-7B LongVA-7B VideoLLaMA2 Kangaroo-8B Qwen2.5-VL-7B (CoT) Qwen2.5-VL-7B-SFT
Builds on: GRPO DeepSeek-R1 Qwen2.5-VL
Closest papers by meaning
- Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.93)
- VideoRFT: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning (NeurIPS 2025 · similarity 0.92)
- ReWatch-R1: Boosting Complex Video Reasoning in Large Vision-Language Models through Agentic Data Synthesis (ICLR 2026 · similarity 0.92)
- OneThinker: All-in-one Reasoning Model for Image and Video (CVPR 2026 · similarity 0.91)
- Scaling RL to Long Videos (NeurIPS 2025 · similarity 0.91)
- Beyond Perceptual Shortcuts: Causal-Inspired Debiasing Optimization for Generalizable Video Reasoning in Lightweight MLLMs (CVPR 2026 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.