Researchers led by Jing Xiong, a lead researcher and the primary paper submitter, have developed MMFormalizer to address the complexities of autoformalization in the physical world. Unlike traditional models that rely solely on text, this framework infers hidden physical constraints from visual elements. The system recursively builds formal propositions from grounded primitives, ensuring every abstraction is backed by visual evidence and anchored in axiomatic grounding. This methodology enables precise machine reasoning within sophisticated mathematical and physical contexts.

To test these capabilities, the team introduced the PhyX-AF benchmark, which includes 115 curated samples spanning classical mechanics, relativity, and quantum physics. The benchmark evaluates a model's ability to translate visual and textual data into verifiable formal logic. Evaluations of frontier models like GPT-5 and Gemini-3-Pro indicate that while these systems demonstrate high semantic accuracy and physical reasoning, they still struggle with complex geometric interpretations.

MMFormalizer stands as the first multimodal autoformalization method capable of processing advanced physics derived from the Hamiltonian. By bridging the gap between raw visual perception and machine-verifiable logic, the framework provides a pathway for AI to solve high-level scientific problems. This development marks a significant shift toward AI systems that can independently interpret and formalize the laws of the physical universe through direct observation.

Researchers led by Jing Xiong, a lead researcher and the primary paper submitter, have developed MMFormalizer to address the complexities of autoformalization in the physical world. Unlike traditional models that rely solely on text, this framework infers hidden physical constraints from visual elements. The system recursively builds formal propositions from grounded primitives, ensuring every abstraction is backed by visual evidence and anchored in axiomatic grounding. This methodology enables precise machine reasoning within sophisticated mathematical and physical contexts.

To test these capabilities, the team introduced the PhyX-AF benchmark, which includes 115 curated samples spanning classical mechanics, relativity, and quantum physics. The benchmark evaluates a model's ability to translate visual and textual data into verifiable formal logic. Evaluations of frontier models like GPT-5 and Gemini-3-Pro indicate that while these systems demonstrate high semantic accuracy and physical reasoning, they still struggle with complex geometric interpretations.

MMFormalizer stands as the first multimodal autoformalization method capable of processing advanced physics derived from the Hamiltonian. By bridging the gap between raw visual perception and machine-verifiable logic, the framework provides a pathway for AI to solve high-level scientific problems. This development marks a significant shift toward AI systems that can independently interpret and formalize the laws of the physical universe through direct observation.