Research map / Video / Video-language understanding
Video-CoE: Reinforcing Video Event Prediction via Chain of Events
Qile Su, Jing Tang, Rui Chen, Lei Sun, Xiangxiang Chu
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 |
|---|---|---|---|
| replace NEP GRPO with CoE-GRPO | Qwen2.5-VL-7B · FutureBench | AVG accuracy: 67.28 -> 75.00 | Improved |
| replace NEP SFT with CoE-SFT | Qwen2.5-VL-3B · FutureBench | AVG accuracy: 57.86 -> 63.60 | Improved |
| replace NEP SFT with CoE-SFT | Qwen2.5-VL-7B · FutureBench | AVG accuracy: 64.39 -> 65.72 | Improved |
| add similarity reward rs (ablation) | CoE-GRPO · FutureBench | AVG accuracy: 72.00 -> 74.61 | Improved |
| increase GRPO group size from 2 to 8 (ablation) | CoE-GRPO · FutureBench | AVG accuracy: 60.61 -> 74.61 -> 77.20 | Improved |
| set event chain length L=3 vs 1 or 5 (ablation) | CoE-GRPO · FutureBench | AVG accuracy: L=1 73.90, L=3 74.61, L=5 71.40 | Mixed |
| prompt-guided detailed video description to boost visual attention (ablation) | Qwen2.5-VL-7B · FutureBench | AVG accuracy: 45.74 vs CoE 75.00 (base 52.94) | Hurt |
| constant bias added to visual-token attention at inference (ablation) | Qwen2.5-VL-7B · FutureBench | AVG accuracy: 52.57 vs CoE 75.00 (base 52.94) | Hurt |
| swap similarity model VideoCLIP-XL for ViCLIP or CLIP-large (ablation) | CoE-GRPO · FutureBench | AVG accuracy: 74.61 vs 73.01 / 74.24 | No clear effect |
| CoE-SFT/CoE-GRPO training vs vanilla SFT on visual attention | Qwen2.5-VL-7B · visual attention | improvement rate (%): SFT -3.33 vs CoE-GRPO +9.20, CoE-SFT +15.11 | Improved |
| CoE-GRPO vs CoE-SFT under open-set judge | Qwen2.5-VL-7B · judge win rate | win rate (%): 38.13 -> 32.42 | Hurt |
Closest papers by meaning
- Black Swan: Abductive and Defeasible Video Reasoning in Unpredictable Events (CVPR 2025 · similarity 0.87)
- VideoReasonBench: Can MLLMs Perform Vision-Centric Complex Video Reasoning? (ICLR 2026 · similarity 0.87)
- EventLens: Event-Structure Reinforcement Learning for Video Understanding (NeurIPS 2026 · similarity 0.86)
- MVBench: A Comprehensive Multi-modal Video Understanding Benchmark (CVPR 2024 · similarity 0.86)
- VidEvent: A Large Dataset for Understanding Dynamic Evolution of Events in Videos (AAAI 2025 · similarity 0.86)
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · similarity 0.86)
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.
