aiqbench analysis
This study confirms that intuitive physics knowledge is reliably encoded in pretrained video foundation models, with performance peaking at intermediate-to-late layers and varying significantly by architecture. V-JEPA demonstrates superior capability in capturing temporal dynamics compared to masked reconstruction and diffusion-based approaches, while disrupting frame order substantially degrades model performance. These findings highlight that the accessibility of physical reasoning depends critically on the pretraining paradigm and the specific readout mechanism employed.
Do Video Foundation Models Understand Intuitive Physics? A Layerwise Probing Analysis
arXiv:2606.09646v1 Announce Type: cross Abstract: We study whether pretrained video foundation models encode intuitive-physics information in their frozen representations, and how this information varies across model families, layers, and probe types. Using frozen-feature probing on IntPhys2 and Minimal Video Pairs (MVP), we compare predictive joint-embedding models (V-JEPA), masked reconstruction models (VideoMAE), and a diffusion-based video generator (LTX-Video). V-JEPA achieves the strongest overall results across benchmarks, especially with probes that model temporal dynamics, while VideoMAE remains competitive and LTX-Video recovers weaker but non-trivial signal. Layerwise analyses show that physics-relevant information is weakest in early layers and becomes most accessible at intermediate-to-late depth, and temporal controls show that disrupting frame order substantially reduces performance, especially on MVP. Together, these results suggest that intuitive-physics knowledge emerges reliably in pretrained video representations, but its accessibility depends strongly on pretraining paradigm, representational depth, and readout mechanism.
Title:Do Video Foundation Models Understand Intuitive Physics? A Layerwise Probing Analysis
Abstract:We study whether pretrained video foundation models encode intuitive-physics information in their frozen representations, and how this information varies across model families, layers, and probe types. Using frozen-feature probing on IntPhys2 and Minimal Video Pairs (MVP), we compare predictive joint-embedding models (V-JEPA), masked reconstruction models (VideoMAE), and a diffusion-based video generator (LTX-Video). V-JEPA achieves the strongest overall results across benchmarks, especially with probes that model temporal dynamics, while VideoMAE remains competitive and LTX-Video recovers weaker but non-trivial signal. Layerwise analyses show that physics-relevant information is weakest in early layers and becomes most accessible at intermediate-to-late depth, and temporal controls show that disrupting frame order substantially reduces performance, especially on MVP. Together, these results suggest that intuitive-physics knowledge emerges reliably in pretrained video representations, but its accessibility depends strongly on pretraining paradigm, representational depth, and readout mechanism.
Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv:2606.09646 [cs.CV]
(or arXiv:2606.09646v1 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2606.09646 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 250 words
Related coverage
How Do Video Foundation Models Encode Intuitive Physics? Probing Across Pretraining Paradigms
arXiv · Sep 16, 2026
Game Design Creativity evals
LessWrong · Oct 12, 2026
Import AI 476: Creativity and RSI; worse-than-China open weight models; AI goes to space
Import AI · Oct 12, 2026
Runway’s WorldPrompt and the Engineering of Real-Time Worlds
Latent Space · Sep 25, 2026