The Perception-Physics Paradox: Probing Scientific Alignment with TC-Bench
For researchers in scientific machine learning and remote sensing, this work highlights a critical failure mode of VFMs and provides a benchmark to evaluate scientific alignment, though the findings are incremental as they confirm known limitations of correlation-based learning.
The paper identifies the Perception-Physics Paradox, where Vision Foundation Models (VFMs) may appear correct on satellite imagery due to visual correlations rather than true scientific understanding. The authors introduce scientific alignment and release TC-Bench, a benchmark for tropical cyclones, showing that current VFMs fail in intense regimes, indicating scientific alignment does not emerge from scaling alone.
While Vision Foundation Models (VFMs) excel at predictive tasks on satellite imagery, their performance can arise from visual correlations rather than underlying structural invariants, making even perception-based out-of-distribution accuracy a poor proxy for scientific utility. As a result, models may look correct without reasoning correctly, a discrepancy we term the Perception-Physics Paradox. To address this gap, we introduce scientific alignment as an implicit objective for representation learning in scientific domains. We study a principled, testable aspect of scientific alignment through structural isomorphism, which requires latent representations to uniquely identify physical systems up to a linear reparameterization. This perspective induces a hierarchy of necessary conditions and yields a systematic probing protocol for physical and causal interpretability. To operationalize this framework, we release TC-Bench, a global, reproducible benchmark dataset with an automated construction pipeline for tropical cyclone research, and show that current VFMs rely on visual shortcuts that collapse in intense regimes, indicating that scientific alignment does not arise as a natural byproduct of scaling alone.