Research map / Video / Video-language understanding
Video-as-Answer: Predict and Generate Next Video Event with Joint-GRPO
Junhao Cheng, Liang Hou, Xin Tao, Jing Liao
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 Joint-GRPO post-training after SFT | VANS · Procedural VNEP | ROUGE-L: 0.2812 -> 0.3631 | Improved |
| Add Joint-GRPO post-training after SFT | VANS · Procedural VNEP | CLIP-V: 0.7655 -> 0.8021 | Improved |
| Add Joint-GRPO post-training after SFT | VANS · Predictive VNEP | FVD: 94.12 -> 86.85 | Improved |
| Optimize VLM and VDM jointly instead of GRPO on each separately (ablation) | VANS · Procedural VNEP | FVD: 83.14 -> 78.32 | Improved |
| Use two-stage Joint-GRPO instead of all-in-one (ablation) | VANS · Procedural VNEP | FVD: 81.01 -> 78.32 | Improved |
| Add Stage 2 VDM adaptation after Stage 1 (ablation) | VANS · Procedural VNEP | CLIP-V: 0.7803 -> 0.8021 | Improved |
| Add text fidelity reward r_t1 in Stage 1 (ablation) | VANS · Procedural VNEP | ROUGE-L: 0.3498 -> 0.3631 | Improved |
| Add video fidelity reward r_v1 in Stage 1 (ablation) | VANS · Procedural VNEP | CLIP-V: 0.7668 -> 0.7803 | Improved |
| Add video fidelity reward r_v1 in Stage 1 (ablation) | VANS · Procedural VNEP | ROUGE-L: 0.3625 -> 0.3631 | No clear effect |
| Add semantic alignment reward r_c2 in Stage 2 (ablation) | VANS · Procedural VNEP | CLIP-T: 0.3673 -> 0.3824 | Improved |
| Add video fidelity reward r_v2 in Stage 2 (ablation) | VANS · Procedural VNEP | FVD: 79.76 -> 78.32 | Improved |
| Apply GRPO only to the VDM (ablation) | VANS · Procedural VNEP | CLIP-T: 0.3202 -> 0.3013 | Hurt |
Closest papers by meaning
- VG-TVP: Multimodal Procedural Planning via Visually Grounded Text-Video Prompting (AAAI 2025 · similarity 0.87)
- VEDIT: Latent Prediction Architecture For Procedural Video Representation Learning (ICLR 2025 · similarity 0.86)
- Vript: A Video Is Worth Thousands of Words (NeurIPS 2024 · similarity 0.86)
- Position: Video as the New Language for Real-World Decision Making (ICML 2024 · similarity 0.86)
- VidEvent: A Large Dataset for Understanding Dynamic Evolution of Events in Videos (AAAI 2025 · similarity 0.85)
- VSTAR: Generative Temporal Nursing for Longer Dynamic Video Synthesis (ICLR 2025 · similarity 0.85)
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.
