Research map / Video / Video-language understanding
ReasonAct: Progressive Training for Fine-Grained Video Reasoning in Small Models
Jiaxin Liu, Zhaolu Kang
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 Stage 1 text-only reasoning (FRE) before video SFT (ablation) | ReasonAct · HMDB51 | accuracy (%): 51.2 -> 59.4 (with V-SFT, vs direct V-SFT) | Improved |
| add FRE stage (ablation) | Qwen2.5-VL-3B · HMDB51 | accuracy (%): 49.3 -> 53.1 | Improved |
| add video CoT fine-tuning (V-SFT) after FRE (ablation) | Qwen2.5-VL-3B · HMDB51 | accuracy (%): 53.1 -> 59.4 | Improved |
| add temporal-aware RL (T-GRPO) after FRE and V-SFT (ablation) | ReasonAct · HMDB51 | accuracy (%): 59.4 -> 67.2 | Improved |
| add temporal consistency reward (ablation) | T-GRPO · HMDB51 | accuracy (%): 61.8 -> 67.2 | Improved |
| add sub-action reward (ablation) | T-GRPO · HMDB51 | accuracy (%): 64.7 -> 67.2 | Improved |
| add format reward (ablation) | T-GRPO · HMDB51 | accuracy (%): 66.1 -> 67.2 | Improved |
| use full sub-action library instead of 25% (ablation) | T-GRPO · HMDB51 | accuracy (%): 65.4 -> 67.2 | Improved |
| grow sub-action library from 75% to 100% (ablation) | T-GRPO · Kinetics-400 | accuracy (%): 78.8 -> 78.9 | No clear effect |
| apply T-GRPO RL directly without FRE and V-SFT (ablation) | Qwen2.5-VL-3B · UCF-101 | accuracy (%): 78.3 -> 84.1 | Improved |
| ReasonAct vs LLaVA-Video | ReasonAct · Kinetics-400 | accuracy (%): 72.1 -> 78.9 | Improved |
Closest papers by meaning
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.92)
- Find, Fix, Reason: Context Repair for Video Reasoning (ICML 2026 · similarity 0.91)
- Beyond Perceptual Shortcuts: Causal-Inspired Debiasing Optimization for Generalizable Video Reasoning in Lightweight MLLMs (CVPR 2026 · similarity 0.90)
- Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.90)
- OneThinker: All-in-one Reasoning Model for Image and Video (CVPR 2026 · similarity 0.90)
- Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning of Vision Language Models (NeurIPS 2025 · 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.
