Research map / Video / Video-language understanding
Reinforce to Learn, Elect to Reason: A Dual Paradigm for Video Reasoning
Songyuan Yang, Weijiang Yu, Jilin Ma, Ziyu Liu, Guijian Tang, Wenjing Yang, Huibin Tan, Nong Xiao
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 frame-sensitive reward (ablation) | RLER-Training · VSIBench | accuracy: 41.0 -> 43.3 | Improved |
| add think-transparency reward (ablation) | RLER-Training · VSIBench | accuracy: 41.9 -> 43.3 | Improved |
| add anti-repetition reward (ablation) | RLER-Training · VSIBench | accuracy: 42.1 -> 43.3 | Improved |
| add RLER-Inference election (ablation) | RLER · VSIBench | accuracy: 41.7 -> 43.3 | Improved |
| use GRPO with rewards instead of SFT (ablation) | RLER-Training · VideoMMMU | accuracy: 49.8 -> 54.2 | Improved |
| add input diversification (K=1 to K=3.1) (ablation) | RLER-Inference · MVBench | accuracy: 69.1 -> 72.9 | Improved |
| use evidence-weighted election instead of majority vote (ablation) | RLER-Inference · LVBench | accuracy: 48.3 -> 50.7 | Improved |
| add outlier trimming (ablation) | RLER-Inference · MVBench | accuracy: 71.4 -> 72.9 | Improved |
| add referee self-check (ablation) | RLER-Inference · MVBench | accuracy: 71.8 -> 72.9 | Improved |
| apply RLER-Training before RLER-Inference (ablation) | RLER-Inference · LVBench | accuracy: 42.5 -> 50.7 | Improved |
| add transparency and redundancy scoring to frame scoring (ablation) | RLER-Inference · MVBench | accuracy: 70.2 -> 72.9 | Improved |
| adaptive election (K=3.1) instead of single sample | RLER · MVBench | accuracy: 69.1 -> 72.9 | Improved |
Closest papers by meaning
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.90)
- Find, Fix, Reason: Context Repair for Video Reasoning (ICML 2026 · similarity 0.90)
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.90)
- Boosting Reasoning in Large Multimodal Models via Activation Replay (CVPR 2026 · similarity 0.90)
- OneThinker: All-in-one Reasoning Model for Image and Video (CVPR 2026 · similarity 0.89)
- ReWatch-R1: Boosting Complex Video Reasoning in Large Vision-Language Models through Agentic Data Synthesis (ICLR 2026 · similarity 0.89)
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.
