Research map / Large language models / Instruction and fine-tuning
VideoRFT: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning
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.
Proposes VideoRFT, a two-stage (SFT on cognitively inspired video CoT data, then GRPO) reinforcement fine-tuning method for Qwen2.5-VL-7B that adds a SigLIP-based semantic-consistency reward for video reasoning.
Problem
Video reasoning in MLLMs lags behind text reasoning; existing video CoT datasets are scarce and low in complexity, and RFT rewards do not enforce that reasoning is grounded in visual evidence.
Approach
A multi-expert CoT pipeline (GPT-4o-mini structured video captions, DeepSeek-R1 blind cognitively-prompted CoT, Qwen2.5-VL cross-modal refinement) builds VideoRFT-CoT-102K for SFT, and a gated semantic-consistency reward computes SigLIP similarity between the video-description span of the reasoning and the video frames during GRPO.
Key results
- VideoRFT-7B reaches 36.8 on VSI-Bench (+5.0 over Qwen2.5-VL-7B, +2.8 over GPT-4o), 68.5 on MMVU and 59.8 on VideoMME, ranking first on four of six benchmarks.
- Removing cross-modal CoT refinement drops VSI-Bench by 2.3, MMVU by 3.7 and VideoMME by 7.0 points.
- The 3B variant beats TinyLLaVA-Video-R1-3B by 8.2 points on MMVU and 11.5 on TempCompass.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply VideoRFT’s SFT-plus-RL training to Qwen2.5-VL-7B | Qwen2.5-VL-7B · VSI-Bench | accuracy: +5.0% | Helped |
| Apply VideoRFT’s SFT-plus-RL training to Qwen2.5-VL-7B | Qwen2.5-VL-7B · MMVU | accuracy: +7.2% | Helped |
| Apply VideoRFT’s SFT-plus-RL training to Qwen2.5-VL-7B | Qwen2.5-VL-7B · VideoMME | accuracy: +7.0% | Helped |
| Add cross-modal CoT refinement to the training data (ablation) | VideoRFT · VSI-Bench | accuracy: -2.3% | Helped |
| Add cross-modal CoT refinement to the training data (ablation) | VideoRFT · MMVU | accuracy: -3.7% | Helped |
| Add cross-modal CoT refinement to the training data (ablation) | VideoRFT · VideoMME | accuracy: -7.0% | Helped |
| Add RL after SFT (ablation) | VideoRFT · Video reasoning and understanding benchmarks | accuracy | Helped |
| Add SFT warm-up before RL (ablation) | VideoRFT · Video reasoning and understanding benchmarks | accuracy | Helped |
| Add semantic-consistency reward to format and accuracy rewards (ablation) | VideoRFT RL · VSI-Bench | accuracy | Helped |
| Add accuracy-gated activation of semantic-consistency reward (ablation) | VideoRFT RL · VSI-Bench | accuracy: +2.2% | Helped |
| Add accuracy-gated activation of semantic-consistency reward (ablation) | VideoRFT RL · MMVU | accuracy: +3.3% | Helped |
| Add accuracy-gated activation of semantic-consistency reward (ablation) | VideoRFT RL · VideoMME | accuracy: +3.5% | Helped |
Surprising findings
- The RL-only variant outperforms the SFT-only variant on four of six benchmarks.
- Adding the semantic-consistency reward ungated lowers VideoMME (56.3 vs 58.5 without it) and MMVU, so the gating matters.
Limitations
- Performance may degrade under rapid motion or severe occlusion.
- CoT quality is bounded by the reasoning LLM used (DeepSeek-R1).
Experimental setup
Datasets: VideoRFT-CoT-102K VideoRFT-RL A-OKVQA ShareGPT4V CLEVR Geometry3K UniGeo AI2D TextVQA HME100k ScienceQA PMC-VQA
Benchmarks: VSI-Bench VideoMMMU MMVU MVBench TempCompass VideoMME
Metrics: accuracy
Baselines: GPT-4o Qwen2.5-VL-7B Video-R1 TinyLLaVA-Video-R1 VideoChat-R1 LLaVA-OneVision-7B LongVA-7B VideoLLaMA2
Builds on: GRPO DeepSeek-R1 Qwen2.5-VL Video-R1 Open-R1
Closest papers by meaning
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.92)
- VideoReasonBench: Can MLLMs Perform Vision-Centric Complex Video Reasoning? (ICLR 2026 · similarity 0.91)
- Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.91)
- Beyond Perceptual Shortcuts: Causal-Inspired Debiasing Optimization for Generalizable Video Reasoning in Lightweight MLLMs (CVPR 2026 · similarity 0.91)
- Reinforcing Structured Chain-of-Thought for Video Understanding (CVPR 2026 · similarity 0.91)
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.90)
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.