Research map / Video / Video segmentation and tracking
GT-SVJ: Generative-Transformer-Based Self-Supervised Video Judge For Efficient Video Reward Modeling
Shivanshu Shekhar, Uttaran Bhattacharya, Raghavendra Addanki, Mehrab Tanjim, Somdeb Sarkhel, Tong Zhang
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 |
|---|---|---|---|
| pretrain a contrastive discriminative model before reward tuning (ablation) | GT-SVJ reward model · GenAI-Bench (w/ties) | preference accuracy (%): 47.62 -> 64.26 | Improved |
| pretrain a contrastive discriminative model before reward tuning (ablation) | GT-SVJ reward model · MonteBench (w/o ties) | preference accuracy (%): 73.59 -> 77.76 | Improved |
| add variable-length video truncation (p=0.25) (ablation) | GT-SVJ discriminative model · GenAI-Bench (w/ties) | preference accuracy (%): 49.36 -> 64.26 | Improved |
| add variable-length video truncation (p=0.25) (ablation) | GT-SVJ discriminative model · MonteBench (w/ties) | preference accuracy (%): 62.79 -> 66.36 | Improved |
| add perturbed videos as negatives alongside generated videos (ablation) | GT-SVJ discriminative model · validation alignment accuracy | accuracy: 2-3% better validation accuracy, faster convergence | Improved |
| add perturbed videos as negatives alongside generated videos (ablation) | GT-SVJ discriminative model · training dynamics | gradient norm / loss | Improved |
| apply LoRA to middle third instead of last third of transformer layers (ablation) | CogVideoX-2B reward model · validation loss and accuracy | validation accuracy | Improved |
| apply LoRA to last third instead of middle third (ablation) | CogVideoX-2B reward model · training speed | training time: about 1.5x faster training, marginal performance decrease | Improved |
| sample perturbations by inverse gradient-decay difficulty instead of uniformly | GT-SVJ discriminative model · negative sampling | training dynamics | No clear effect |
| use a generative video model backbone instead of a VLM as reward model | GT-SVJ · GenAI-Bench (w/ties) | preference accuracy (%): 51.56 -> 64.26 vs VisionReward | Improved |
| use a generative video model backbone instead of a VLM as reward model | GT-SVJ · VideoReward-Bench (w/ties) | preference accuracy (%): 61.26 -> 57.01 vs VideoReward | Hurt |
Closest papers by meaning
- Think, then Score: Decoupled Reasoning and Scoring for Video Reward Modeling (NeurIPS 2026 · similarity 0.89)
- Improving Video Generation with Human Feedback (NeurIPS 2025 · similarity 0.88)
- Video-Bench: Human-Aligned Video Generation Benchmark (CVPR 2025 · similarity 0.87)
- InstructVideo: Instructing Video Diffusion Models with Human Feedback (CVPR 2024 · similarity 0.87)
- Boosting Text-to-Video Generative Model with MLLMs Feedback (NeurIPS 2024 · similarity 0.86)
- VBench: Comprehensive Benchmark Suite for Video Generative Models (CVPR 2024 · 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.
