Title: The Perception-Physics Paradox: Probing Scientific Alignment with TC-Bench

Abstract:

Although Vision Foundation Models (VFMs) demonstrate strong predictive capabilities when analyzing satellite imagery, their success often stems from exploiting visual correlations rather than capturing underlying structural invariants. Consequently, high accuracy on out-of-distribution perception tasks serves as an unreliable indicator of scientific utility. This phenomenon creates a situation where models appear accurate without possessing correct reasoning, a discrepancy the authors label the "Perception-Physics Paradox." To bridge this divide, the study proposes scientific alignment as an implicit objective within representation learning for scientific fields. The research investigates a rigorous, testable dimension of this alignment through the lens of structural isomorphism, which mandates that latent representations must uniquely identify physical systems, allowing only for linear reparameterization. This approach establishes a hierarchy of necessary conditions and facilitates a systematic protocol for assessing physical and causal interpretability. To implement this framework, the authors introduce TC-Bench, an automated, reproducible, and global benchmark dataset tailored for tropical cyclone research. Their findings reveal that existing VFMs depend on visual shortcuts that fail under intense conditions, suggesting that scientific alignment is not an automatic consequence of model scaling.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC