Research map / Video / Video-language understanding
Thinking With Videos: Multimodal Tool-Augmented Reinforcement Learning for Long Video Reasoning
Haoji Zhang, Xin Gu, Jiawen Li, Chixiang Ma, Sule Bai, Chubin Zhang, Bowen Zhang, Zhichao Zhou, Dongliang He, Yansong Tang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DGRPO (difficulty-aware rewards) in place of GRPO (ablation) | VITAL-7B (SFT, think, no tools) · LVR / VidCh / MMMU / Cha average | average score: 50.3 -> 52.1 | Improved |
| add DGRPO in place of GRPO (ablation) | VITAL-7B · LongVideo-Reason | accuracy: 66.0 -> 70.2 | Improved |
| add DGRPO in place of GRPO (ablation) | VITAL-7B · Charades-STA | mIoU: 57.2 -> 57.1 | No clear effect |
| add tool-augmented SFT and DGRPO stages (video clip tool) (ablation) | VITAL-7B · LVR / VidCh / MMMU / Cha average | average score: 52.1 -> 57.1 | Improved |
| add toolbox at inference (ablation) | VITAL-7B · LongVideo-Reason | accuracy: 73.2 -> 79.3 | Improved |
| add toolbox at inference (ablation) | VITAL-7B · VidChapters-7M | R@0.5: 25.8 -> 34.7 | Improved |
| add thinking (reasoning) style to SFT+GRPO (ablation) | VITAL-7B · LVR / VidCh / MMMU / Cha average | average score: 48.3 -> 50.3 | Improved |
| add GRPO after SFT (ablation) | VITAL-7B (think) · LVR / VidCh / MMMU / Cha average | average score: 43.9 -> 50.3 | Improved |
| train on temporal grounding + reasoning VQA + grounded VQA instead of a single task (same 73k size) (ablation) | VITAL-7B · LVR / VidCh / MMMU / Cha average | average score: 45.4 (TG) / 42.4 (RQA) -> 53.9 | Improved |
| scale training data from 73k to 182k (ablation) | VITAL-7B · LVR / VidCh / MMMU / Cha average | average score: 53.9 -> 57.1 | Improved |
| use clip caption or clip QA tool (Gemini-2.5-Pro) instead of video clip tool (ablation) | VITAL-7B · LVR / VidCh / MMMU / Cha average | average score: 57.1 -> 51.8 (caption) / 53.0 (QA) | Hurt |
| add clip caption or clip QA tool zero-shot to GPT-4.1 (ablation) | GPT-4.1 · VidChapters-7M temporal grounding (zero-shot) | mIoU: 28.4 -> 2.0 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ReVisionLLM — ReVisionLLM: Recursive Vision-Language Model for Temporal Grounding in Hour-Long Videos (CVPR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.92)
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.91)
- VideoRFT: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning (NeurIPS 2025 · similarity 0.91)
- VideoEspresso: A Large-Scale Chain-of-Thought Dataset for Fine-Grained Video Reasoning via Core Frame Selection (CVPR 2025 · similarity 0.91)
- Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation (CVPR 2025 · similarity 0.91)
- OneThinker: All-in-one Reasoning Model for Image and Video (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.
