Research map / Video / Video-language understanding
PAI-Bench: A Comprehensive Benchmark For Physical AI
Fengzhe Zhou, Jiannan Huang, Jialuo Li, Deva Ramanan, Humphrey Shi
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 |
|---|---|---|---|
| use multi-signal (All) conditioning instead of the best single control signal (ablation) | Cosmos-Transfer1-7B · PAI-Bench-C | Quality Score: 6.89 -> 9.24 | Improved |
| use multi-signal (All) conditioning instead of the best single control signal (ablation) | Cosmos-Transfer2.5-2B · PAI-Bench-C | Quality Score: 8.77 -> 9.24 | Improved |
| use segmentation maps instead of all signals as control (ablation) | Cosmos-Transfer1-7B · PAI-Bench-C | Mask mIoU: 0.74 -> 0.70 | Hurt |
| use blur control instead of all signals (ablation) | Cosmos-Transfer2.5-2B · PAI-Bench-C | Diversity: 0.13 -> 0.18 | Improved |
| scale Qwen2.5-VL from 7B to 72B | Qwen2.5-VL · PAI-Bench-U | accuracy: 51.0 -> 60.8 | Improved |
| scale Qwen3-VL from 8B to 235B-A22B | Qwen3-VL · PAI-Bench-U | accuracy: 56.8 -> 64.7 | Improved |
| scale Cosmos-Predict2 from 2B to 14B | Cosmos-Predict2 · PAI-Bench-G | Overall: 79.7 -> 80.0 | No clear effect |
| scale MAGI-1 from 4.5B to 24B | MAGI-1 · PAI-Bench-G | Overall: 77.5 -> 79.4 | Improved |
| enable textual thinking mode (ablation) | Qwen3-VL series · PAI-Bench-U | accuracy | Hurt |
| use medium instead of minimal reasoning effort (thinking with visual reasoning) (ablation) | GPT-5 · PAI-Bench-U | accuracy | Improved |
| increase input frames from 0 (text-only) to visual frames (ablation) | MLLMs · PAI-Bench-U | accuracy | Improved |
| increase input frames beyond 8 (ablation) | MLLMs · PAI-Bench-U | accuracy | No clear effect |
Closest papers by meaning
- PhysBench: Benchmarking and Enhancing Vision-Language Models for Physical World Understanding (ICLR 2025 · similarity 0.89)
- QUANTIPHY: A Quantitative Benchmark Evaluating Physical Reasoning Abilities of Vision-Language Models (CVPR 2026 · similarity 0.89)
- PhyCritic: Multimodal Critic Models for Physical AI (CVPR 2026 · similarity 0.87)
- Video-Bench: Human-Aligned Video Generation Benchmark (CVPR 2025 · similarity 0.87)
- VideoHallu: Evaluating and Mitigating Multi-modal Hallucinations on Synthetic Video Understanding (NeurIPS 2025 · similarity 0.87)
- Apple-$\pi$: Benchmarking Thinking with Video Towards Law-Grounded Physical Intelligence (NeurIPS 2026 · 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.
