Research map / Video / Video-language understanding
Reinforcing Video Reasoning with Focused Thinking
Jisheng Dang, Jingze Wu, Teng Wang, Xuanhui Lin, Nannan Zhu, Hongbo Chen, WEISHI ZHENG, Meng Wang, Tat-Seng Chua
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 |
|---|---|---|---|
| train on multi-choice instead of single-choice questions (ablation) | GRPO · CLEVRER counterfactual | accuracy (all): 32.6 -> 41.1 | Improved |
| train on multi-choice instead of single-choice questions (ablation) | GRPO · NExT-GQA source | accuracy (all): 30.7 -> 36.7 | Improved |
| train on multi-choice instead of single-choice questions (ablation) | GRPO · STAR source | accuracy (all): 29.3 -> 29.1 | No clear effect |
| train on multi-choice instead of single-choice (ablation) | TW-GRPO · CLEVRER counterfactual | accuracy (all): 38.9 -> 50.4 | Improved |
| use soft reward instead of fixed reward (ablation) | TW-GRPO · CLEVRER counterfactual | soft accuracy (all): 58.7 -> 62.9 | Improved |
| add token weighting (alpha=0.7) (ablation) | GRPO with soft reward · CLEVRER counterfactual | soft accuracy (multiple choice): 57.6 -> 64.4 | Improved |
| add token weighting (ablation) | GRPO with fixed reward · CLEVRER counterfactual | accuracy (all): 41.1 -> 42.6 | Improved |
| use nonzero token weight alpha=0.7 instead of alpha=0 (ablation) | TW-GRPO · CLEVRER counterfactual | accuracy (all): 48.6 -> 50.4 | Improved |
| raise sampling number from 4 to 8 (ablation) | TW-GRPO · CLEVRER counterfactual | accuracy (all): 38.9 -> 50.4 | Improved |
| raise sampling number from 8 to 12 (ablation) | TW-GRPO · CLEVRER counterfactual | accuracy (all): 50.4 -> 49.1 | Hurt |
| temperature 1.0 instead of 0.5 (ablation) | TW-GRPO · CLEVRER counterfactual | accuracy (all): 46.3 -> 50.4 | Improved |
| temperature 1.0 instead of 1.5 (ablation) | TW-GRPO · CLEVRER counterfactual | accuracy (all): 44.3 -> 50.4 | Improved |
Closest papers by meaning
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.93)
- Find, Fix, Reason: Context Repair for Video Reasoning (ICML 2026 · similarity 0.92)
- Think or Not? Selective Reasoning via Reinforcement Learning for Vision-Language Models (NeurIPS 2025 · similarity 0.92)
- OneThinker: All-in-one Reasoning Model for Image and Video (CVPR 2026 · similarity 0.91)
- DeepVideo-R1: Video Reinforcement Fine-Tuning via Difficulty-aware Regressive GRPO (NeurIPS 2025 · similarity 0.91)
- Reason in the Words You Speak: Idiolectal Paraphrasing Off-Policy Traces for Reasoning Distillation in VideoLLMs (NeurIPS 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.
