Geophysical Evidence-Guided Preference Alignment for Earth-Observation Vision-Language Models under Domain Shift

Earth-observation vision-language models (EO-VLMs) are now expected to describe, questionanswer and localise land-surface content from satellite and aerial imagery, but their alignment stage is still inherited from general-purpose multimodal training, where every token of a preferred answer is treated as equally informative and where the margin that separates a preferred from a rejected answer is a constant. Neither assumption holds over the land surface: the tokens that decide whether a description is correct are a small minority, they are correct only if the imagery supports them, and what counts as evidence changes with biome, sensor, season and ground sampling distance. We present an alignment framework in which the preference signal is priced by physical image evidence. GIST derives token importance from an evidence map built from spectral indices and multi-scale saliency, aligns text spans to image regions through referring expressions, and fuses evidence support with gradient importance into a length-normalised token weight. GRACE replaces the constant margin with the difference of evidence-consistency scores between the preferred and rejected answers, adds a hallucination penalty on unsupported entity mentions, and trims preference pairs whose consistency gap carries high posterior variance. TERRA conditions the importance head on a descriptor vector (biome, sensor, season, resolution), normalises importance within descriptor groups so that order rather than scale transfers, and re-estimates evidence on unlabelled target imagery with an elastic anchoring term.