GEOPHYS: The Geometry of Physical Plausibility

Researchers introduce GEOPHYS, a method that identifies physically implausible events in videos by analyzing geometric properties of image encoder embeddings, achieving 98.3% accuracy on physics-violation detection while being significantly faster and more efficient than existing LLM-based approaches.

GEOPHYS addresses a fundamental limitation in current machine learning systems: the inability to quickly assess physical plausibility in visual content. While humans instinctively recognize physically impossible events within milliseconds, existing AI solutions rely on expensive multimodal large language models or require specialized training modifications. The research demonstrates that frozen image encoders already capture implicit signals about physical plausibility through five measurable geometric properties of their embeddings, eliminating the need for additional computational overhead.

The breakthrough stems from recognizing that physical understanding may be an emergent property of visual encoders trained on large-scale datasets, rather than requiring explicit physical reasoning modules. By analyzing temporal feature geometry rather than semantic content, GEOPHYS achieves state-of-the-art results that dramatically outperform cutting-edge models including V-JEPA 2, GPT-4o, and Gemini, which perform near chance on physics-violation detection tasks.

For the AI industry, GEOPHYS has immediate practical applications in video generation and verification. When deployed as a verifier for physics-aligned video generation, it improves MAGI-1 24B's performance from 50.01% to 64.50% on PhysicsIQ benchmarks while consuming 4.65x less memory and running 1.5x faster than alternative approaches. This efficiency gain is crucial for scaling video generation systems in production environments.

The findings suggest that physical reasoning may not require specialized architectures or reasoning modules but can leverage existing vision infrastructure more effectively. Future work may explore whether similar geometric principles apply to other forms of semantic understanding, potentially enabling faster and more efficient verification systems across multiple domains.

• →GEOPHYS achieves 98.3% accuracy on physics-violation detection using only geometric properties of image encoder embeddings
• →The method outperforms GPT-4o, Gemini, and modern video diffusion models while consuming significantly fewer computational resources
• →Physical plausibility understanding emerges implicitly from frozen image encoders without requiring specialized training or external LLM judges
• →GEOPHYS improves video generation alignment verification by 24.5% while reducing memory consumption by 4.65x compared to world-model approaches
• →The research demonstrates that efficient physical reasoning in AI may require leveraging emergent geometric properties rather than building specialized reasoning modules