Research map / Large language models / Reasoning and chain of thought
Wan-R1: Verifiable-Reinforcement Learning for Generalizable Video Reasoning
Ming Liu, Yunbei Zhang, Shilong Liu, Liwen Wang, Wensheng Zhang
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 Flow-GRPO RL with verifiable rewards after SFT | Wan-SFT · VR-Bench Regular Maze | exact match: 33.3 -> 61.1 | Improved |
| add Flow-GRPO RL with verifiable rewards after SFT | Wan-SFT · VR-Bench Trapfield | exact match: 38.9 -> 90.3 | Improved |
| add Flow-GRPO RL with verifiable rewards after SFT | Wan-SFT · VR-Bench 3D Maze | exact match: 65.3 -> 94.4 | Improved |
| add Flow-GRPO RL with verifiable rewards after SFT | Wan-SFT · VR-Bench Irregular Maze | exact match: 56.9 -> 47.9 | Hurt |
| add RL to SFT model, trained on Easy only, tested on Hard | Wan-SFT · VR-Bench 3D Maze Hard | precision rate: 58.7 -> 96.2 | Improved |
| add RL to SFT, Irregular Maze Hard generalization | Wan-SFT · VR-Bench Irregular Maze Hard | precision rate: 68.4 -> 51.6 | Hurt |
| raise KL coefficient from 0 to 0.1 (ablation) | Wan-R1 · VR-Bench Regular Maze | step deviation: 6.7 -> 3.7 | Improved |
| raise KL coefficient from 0 to 0.1 (ablation) | Wan-R1 · VR-Bench Regular Maze | exact match: 62.5 -> 66.7 | Improved |
| reduce training denoising steps from 50 to 30 (ablation) | Wan-R1 · VR-Bench Regular Maze | exact match: 62.5 -> 61.1 | No clear effect |
| reduce training denoising steps from 50 to 5 (ablation) | Wan-R1 · VR-Bench Regular Maze | exact match: 62.5 -> 59.7 | Hurt |
| use Qwen2.5-VL as learned reward instead of verifiable reward (ablation) | Wan-R1 · VR-Bench mazes | reward hacking | Hurt |
| use sparse reward only (SR-Only or exact match) (ablation) | Wan-R1 · VR-Bench mazes | exact match | Hurt |
Later papers that compared with this method
1 later accepted paper reports Wan-R1 as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- ReaForest: Fostering Generative Video Reasoning for Spatial Planning (ICML 2026 · 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
- From Verifiable Dot to Reward Chain: Harnessing Verifiable Reference-based Rewards for Reinforcement Learning of Open-ended Generation (ICLR 2026 · similarity 0.90)
- SOLE-R1: Video-Language Reasoning as the Sole Reward for On-Robot Reinforcement Learning (NeurIPS 2026 · similarity 0.90)
- The Reasoning Boundary Paradox: How Reinforcement Learning Constrains Language Models (NeurIPS 2026 · similarity 0.89)
- Synthesizing Multimodal Verifiable Game Data to Boost VLMs' General Reasoning (ICLR 2026 · similarity 0.88)
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.88)
- ReAgent-V: A Reward-Driven Multi-Agent Framework for Video Understanding (NeurIPS 2025 · similarity 0.88)
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.
