Research map / Video / Video generation
Building a Precise Video Language with Human-AI Oversight
Zhiqiu Lin, Siyuan Cen, Chancharik Mitra, Isaac Li, Yuhan Huang, Yu Tong Tiffany Ling, Hewei Wang, Irene Pi, Shihang Zhu, Yili Han, Yilun Du, Deva Ramanan
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 explicit preference and critique signals (full data) to SFT (ablation) | Qwen3-VL-8B-Instruct SFT · CHAI benchmark reward modeling | accuracy: 50.9 -> 89.8 | Improved |
| add explicit preference and critique signals (full data) to SFT (ablation) | Qwen3-VL-8B-Instruct SFT · CHAI benchmark critique generation | BLEU-4: 5.5 -> 41.7 | Improved |
| add explicit preference and critique signals (full data) to SFT (ablation) | Qwen3-VL-8B-Instruct SFT · CHAI benchmark caption generation | BLEU-4: 12.0 -> 18.2 | Improved |
| add explicit preference and critique signals to DPO (ablation) | Qwen3-VL-8B-Instruct DPO · CHAI benchmark caption generation | BLEU-4: 9.0 -> 15.8 | Improved |
| add explicit preference and critique signals to RLHF-V (ablation) | Qwen3-VL-8B-Instruct RLHF-V · CHAI benchmark reward modeling | accuracy: 50.0 -> 81.0 | Improved |
| use SFT instead of DPO on full data | Qwen3-VL-8B-Instruct · CHAI benchmark caption generation | BLEU-4: 15.8 -> 18.2 | Improved |
| add second-stage quality control to critiques (ablation) | SFT with human critiques · CHAI benchmark reward modeling | accuracy: 73.1 -> 89.8 | Improved |
| add second-stage quality control to critiques (ablation) | SFT with human critiques · CHAI benchmark critique generation | BLEU-4: 35.7 -> 41.7 | Improved |
| ensure critique precision (avoid inaccurate critiques) (ablation) | SFT with human critiques · CHAI benchmark reward modeling | accuracy: 47.1 -> 89.8 | Improved |
| ensure critique recall (avoid incomplete critiques) (ablation) | SFT with human critiques · CHAI benchmark reward modeling | accuracy: 56.6 -> 89.8 | Improved |
| ensure critique constructiveness (avoid unhelpful critiques) (ablation) | SFT with human critiques · CHAI benchmark critique generation | BLEU-4: 32.9 -> 41.7 | Improved |
| use human expert critiques instead of Gemini-2.5-Pro critiques (ablation) | SFT · CHAI benchmark reward modeling | accuracy: 62.0 -> 89.8 | Improved |
Closest papers by meaning
- Distilling Vision-Language Models on Millions of Videos (CVPR 2024 · similarity 0.90)
- Auteur: Language-Driven Cinematographic Framing for Human-Centric Video Generation (NeurIPS 2026 · similarity 0.88)
- CAPability: A Comprehensive Visual Caption Benchmark for Evaluating Both Correctness and Thoroughness (NeurIPS 2025 · similarity 0.88)
- Progress-Aware Video Frame Captioning (CVPR 2025 · similarity 0.88)
- Can Text-to-Video Generation help Video-Language Alignment? (CVPR 2025 · similarity 0.88)
- ShotBench: Expert-Level Cinematic Understanding in Vision-Language Models (NeurIPS 2025 · similarity 0.88)
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.
