Research map / Video / Video-language understanding
ReaSon: Reinforced Causal Search with Information Bottleneck for Video Understanding
Yuan Zhou, Litao Hua, Shilong Jin, Wentao Huang, Haoran Duan
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 cycle consistency reward to answer reward (ablation) | ReaSon (LLaVA-Video-7B, 8 frames) · NExT-QA val | Avg accuracy: 80.1 -> 80.5 | Improved |
| add counterfactual reward on top of answer and cycle rewards (ablation) | ReaSon (LLaVA-Video-7B, 8 frames) · NExT-QA val | Avg accuracy: 80.5 -> 81.4 | Improved |
| add cycle consistency reward to answer reward (ablation) | ReaSon (LLaVA-Video-7B, 8 frames) · EgoSchema subset | accuracy: 66.0 -> 68.2 | Improved |
| add counterfactual reward on top of answer and cycle rewards (ablation) | ReaSon (LLaVA-Video-7B, 8 frames) · EgoSchema subset | accuracy: 68.2 -> 69.0 | Improved |
| add ReaSon frame selection to GPT-4o | GPT-4o (8 frames) · NExT-QA val | Avg accuracy: 72.0 -> 77.6 | Improved |
| add ReaSon frame selection to Qwen2.5-VL-7B | Qwen2.5-VL-7B (8 frames) · EgoSchema subset | accuracy: 65.8 -> 68.0 | Improved |
| add ReaSon frame selection to LLaVA-Video-7B | LLaVA-Video-7B (8 frames) · EgoSchema subset | accuracy: 65.2 -> 69.0 | Improved |
| add ReaSon frame selection to GPT-4o | GPT-4o (8 frames) · Video-MME (short videos) | accuracy: 55.7 -> 65.9 | Improved |
| use ReaSon instead of T* with GPT-4o on long videos | GPT-4o (32 frames) · Video-MME long videos | accuracy: 59.3 -> 58.2 | Hurt |
| add ReaSon to GPT-4o | GPT-4o (32 frames) · Video-MME (overall) | Overall accuracy: 61.8 -> 66.4 | Improved |
| use ReaSon with LLaVA-Video-7B instead of uniform GPT-4o at 8 frames | LLaVA-Video-7B (8 frames) · Video-MME (overall) | Overall accuracy: 53.3 vs T* 56.5 (GPT-4o) | Hurt |
Closest papers by meaning
- Efficient Frame Selection for Long Video Understanding via Reinforcement Learning (CVPR 2026 · similarity 0.90)
- Logic-in-Frames: Dynamic Keyframe Search via Visual Semantic-Logical Verification for Long Video Understanding (NeurIPS 2025 · similarity 0.89)
- Perception Without Engagement: Dissecting the Causal Discovery Deficit in LMMs (NeurIPS 2026 · similarity 0.88)
- Find, Fix, Reason: Context Repair for Video Reasoning (ICML 2026 · similarity 0.88)
- Beyond Perceptual Shortcuts: Causal-Inspired Debiasing Optimization for Generalizable Video Reasoning in Lightweight MLLMs (CVPR 2026 · similarity 0.87)
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · 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.
