A new paper posted on arXiv asks whether vision models can learn physics from what they hear. The authors note that physical understanding is crucial for decision making, yet vision models often fail to infer properties like damping—the way materials dissipate energy and govern how they behave.
The paper argues that audio is a natural source for this information, because an object's acoustic response directly encodes its damping structure. The study's approach is to probe and distill that damping structure, presumably using audio as a teacher signal to improve vision models, though the abstract does not detail specific results.
Because the source is only an abstract, the article cannot confirm findings or methodology beyond the stated motivation and approach. As with all preprints, the work has not yet undergone peer review, and its conclusions should be treated as preliminary.