Research map / Video / Video-language understanding
VideoJudge: Bootstrapping Enables Scalable Supervision of MLLM-as-a-Judge for Video Understanding
Abdul Waheed, Zhen Wu, Dareen Safar Alharthi, Seungone Kim, Bhiksha Raj
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 judge on bootstrapped data (VideoJudge-3B vs base Qwen2.5-VL-3B) | Qwen2.5-VL-3B · VideoJudgeLLaVA-MetaEval | Spearman: 0.63 -> 0.82 | Improved |
| train judge on bootstrapped data (VideoJudge-7B vs base Qwen2.5-VL-7B) | Qwen2.5-VL-7B · VideoJudgeVCG-MetaEval | Spearman: 0.65 -> 0.74 | Improved |
| train judge on bootstrapped data (VideoJudge-3B) | Qwen2.5-VL-3B · VideoJudgeVCG-MetaEval | Spearman: 0.51 -> 0.59 | Mixed |
| train judge on bootstrapped data | Qwen2.5-VL-3B · VATEX-Eval | RMSE: 2.27 -> 1.33 | Improved |
| enable thinking mode for text-only Qwen3 judges (ablation) | Qwen3 (0.6B-4B) · VideoJudgeLLaVA-MetaEval | Spearman: 0.6B: 0.64 -> 0.64; 4B: 0.72 -> 0.68 | Mixed |
| use text-only LLM judge with video descriptions instead of video MLLM judge (ablation) | Qwen3 vs Qwen2.5-VL judges · Meta-evaluation benchmarks | judging performance | Hurt |
| train judge to generate instance-specific rubrics at test time | Qwen2.5-VL-3B · 1000 sampled examples (LLaVA and VCG MetaEval) | MAE: 1.15 -> 0.59 | Improved |
| add feedback generation to pairwise judging (ablation) | Qwen2.5-VL-3B (base) · VideoJudge-Pairwise | accuracy: 75.00 -> 82.60 | Improved |
| add feedback generation to pairwise judging (ablation) | VideoJudge-7B · VideoJudge-Pairwise | accuracy: 98.60 -> 95.60 | Hurt |
| increase training max frames up to about 240 (ablation) | VideoJudge · Pointwise meta-evaluation | correlation: correlation exceeds 0.7 up to ~240 frames | Improved |
| increase evaluation max frames beyond 120 (ablation) | VideoJudge · Pointwise meta-evaluation | correlation | No clear effect |
| raise decoding temperature from 0.0 to 1.0 (ablation) | Qwen2.5-VL-3B base · Pointwise meta-evaluation | Spearman: 0.56 -> 0.42 | Hurt |
Closest papers by meaning
- Needle In A Video Haystack: A Scalable Synthetic Evaluator for Video MLLMs (ICLR 2025 · similarity 0.93)
- MMBench-Video: A Long-Form Multi-Shot Benchmark for Holistic Video Understanding (NeurIPS 2024 · similarity 0.91)
- MVBench: A Comprehensive Multi-modal Video Understanding Benchmark (CVPR 2024 · similarity 0.90)
- Leveraging Textual Memory and Key Frame Reasoning for Full Video Understanding Using Off-the-Shelf LLMs and VLMs (Student Abstract) (AAAI 2025 · similarity 0.90)
- VideoReasonBench: Can MLLMs Perform Vision-Centric Complex Video Reasoning? (ICLR 2026 · similarity 0.89)
- MVU-Eval: Towards Multi-Video Understanding Evaluation for Multimodal LLMs (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.
