Research map / Video / Video-language understanding
EVA: Efficient Reinforcement Learning for End-to-End Video Agent
Yaolun Zhang, Ruohui Wang, Jiahao Wang, Yepeng Tang, Xuanyu Zheng, Haonan Duan, Hao Lu, Hanming Deng, Lewei Lu
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 KTO stage after SFT (ablation) | EVA · LongVideoBench | accuracy (%): 49.9 -> 53.2 | Improved |
| add KTO stage after SFT (ablation) | EVA · MLVU | accuracy (%): 52.3 -> 57.4 | Improved |
| add KTO stage after SFT (ablation) | EVA · LVBench | accuracy (%): 26.5 -> 36.0 | Improved |
| add KTO stage after SFT (ablation) | EVA · VideoMME-Overall | accuracy (%): 56.0 -> 56.5 | No clear effect |
| add KTO stage after SFT (ablation) | EVA · VideoMME-Long | accuracy (%): 45.8 -> 45.1 | Hurt |
| add GRPO stage after KTO (ablation) | EVA · LongVideoBench | accuracy (%): 53.2 -> 55.0 | Improved |
| add GRPO stage after KTO (ablation) | EVA · MLVU | accuracy (%): 57.4 -> 68.3 | Improved |
| add GRPO stage after KTO (ablation) | EVA · VideoMME-Overall | accuracy (%): 56.5 -> 60.2 | Improved |
| add GRPO stage after KTO (ablation) | EVA · LVBench | accuracy (%): 36.0 -> 43.3 | Improved |
| add GRPO stage after KTO (ablation) | EVA · Video-Holmes (zero-shot) | overall accuracy (%): 32.9 -> 37.2 | Improved |
| mix open-ended QA with multiple-choice QA in GRPO data (ablation) | EVA-GRPO · VideoMME | accuracy | Improved |
| use EVA agent instead of uniform 32-frame sampling | Qwen2.5-VL-7B · LSDBench | accuracy (%): 49.2 -> 51.0 | Improved |
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.
- FrameThinker — FrameThinker: Learning to Think with Long Videos via Multi-Turn Frame Spotlighting (ICLR 2026) · better on 3 of 3 shared results
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
- ReAgent-V: A Reward-Driven Multi-Agent Framework for Video Understanding (NeurIPS 2025 · similarity 0.89)
- Videos are Sample-Efficient Supervisions: Behavior Cloning from Videos via Latent Representations (NeurIPS 2025 · similarity 0.88)
- Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.88)
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.88)
- ReWatch-R1: Boosting Complex Video Reasoning in Large Vision-Language Models through Agentic Data Synthesis (ICLR 2026 · similarity 0.88)
- PyVision-RL: Forging Open Agentic Vision Models via RL (ICML 2026 · similarity 0.87)
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.
