Research map / Vision-language and multimodal / Multimodal LLMs
Human-Level Accuracy, Non-Human Strategies: Revealing Model-Human Divergence in Video Physical Reasoning
Fanhong Li, Shurui Zheng, Yinzi Yinzi, Junbo Cui, Lei Ji, Jia Liu
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 object markers to the videos (ablation) | V-JEPA2 · Physion overall (frame step 5) | accuracy (%): 69.2 -> 71.6 | Improved |
| add object markers to the videos (ablation) | V-JEPA2 · Physion containment | accuracy (%): 68.7 -> 78.7 | Improved |
| add object markers to the videos (ablation) | V-JEPA2 · Physion collision | accuracy (%): 71.3 -> 78.7 | Improved |
| add object markers to the videos (ablation) | V-JEPA2 · Physion roll/slide | accuracy (%): 84.0 -> 80.0 | Hurt |
| use token-level probe instead of mean-pooled attentive probe (ablation) | V-JEPA2 · Physion (frame step 5) | accuracy (%): 62.50% -> 67.42% | Improved |
| preserve spatial tokens (CrossAttn-GRU) instead of spatial mean pooling (GRU-Pool) (ablation) | V-JEPA2 probe · Physion roll/slide | accuracy (%): +15.3% | Improved |
| use V-JEPA2 backbone instead of VideoMAEv2 | ViT-L probe · Physion | accuracy (%): 67.65 -> 73.16 | Improved |
| use V-JEPA2 backbone instead of DINOv2 | ViT-L probe · Physion | accuracy (%): 66.90 -> 73.16 | Improved |
| require future prediction (target not yet in contact at cutoff) | V-JEPA2 · Physion model-human disagreement | disagreement (%): 10.5% -> 26.8% | Hurt |
| majority-vote ensemble of all four models instead of V-JEPA2 alone (ablation) | V-JEPA2, RVM, VideoMAEv2, DINOv2 · Physion (10 seeds) | accuracy (%): 73.3 -> 71.8 | Hurt |
| select probe head by test oracle instead of train accuracy (ablation) | V-JEPA2 · Physion | accuracy (pp): 0.36pp | No clear effect |
| use a 5.75 frame step (1.5s) instead of 5 (1.17s) (ablation) | V-JEPA2 · Physion overall | accuracy (%): 71.6 -> 73.2 | Improved |
Closest papers by meaning
- HumanPCR: Probing MLLM Capabilities in Diverse Human-Centric Scenes (ICLR 2026 · similarity 0.89)
- QUANTIPHY: A Quantitative Benchmark Evaluating Physical Reasoning Abilities of Vision-Language Models (CVPR 2026 · similarity 0.88)
- VideoHallu: Evaluating and Mitigating Multi-modal Hallucinations on Synthetic Video Understanding (NeurIPS 2025 · similarity 0.87)
- SciVideoBench: Benchmarking Scientific Video Reasoning in Large Multimodal Models (ICML 2026 · similarity 0.87)
- TOMATO: Assessing Visual Temporal Reasoning Capabilities in Multimodal Foundation Models (ICLR 2025 · similarity 0.87)
- Black Swan: Abductive and Defeasible Video Reasoning in Unpredictable Events (CVPR 2025 · similarity 0.87)
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.
