Canonical Color as a Lens into Concept Decodability in Vision Encoders and VLMs

Vision-language models can describe an image with remarkable accuracy, yet a more fundamental question remains unanswered: what visual information actually drives their answers? In this work, we investigate this question through causal tracing, and we observe that highly causal vision tokens often lie outside the target region. Extending the analysis to larger vision-language models reveals a similar pattern across models and corruption settings, suggesting that strong multimodal performance does not necessarily imply spatially localized causal representations. We further investigate: can these models preserve visual structure when appearance cues are removed? and find that visual cues are exploited to understand visual structures. Together, our experiments expose a gap between seeing, using, and reasoning over visual structure, and provide a causal framework for studying how visual information is transformed, preserved, and ultimately used by modern vision-language models.

The standard way to read latent knowledge out of a model, a linear probe confirmed by a steering recovery, can systematically overstate what a vision-language model (VLM) actually grounds in the image. We show this on spatial reasoning, where the error is invisible to both probing and steering yet exposed by a one-line causal control: replacing the image with a gray blank. Probes decode the within-axis answer at 73--97% across axes, and a training-free projection lifts a near-chance axis from 59% to 79%, exactly the signature of unlocking latent knowledge. The blank-image arbiter refutes it, revealing three grounding regimes that probing conflates: an axis can be grounded (vision-dependent, correct), a prior (vision-independent, with its decode and its apparent recovery a directional default rather than perception), or, surprisingly, inverted: decodable, causally controllable, but deployed with the wrong sign, so the model scores below chance and the error requires looking. The taxonomy holds across the studied VLMs: in fourteen models spanning six language-model families and 2B--27B, horizontal is grounded, vertical is a prior, and depth is inverted, with the inversion emerging at scale within families. The decode-versus-deploy inversion replicates on seven of eight models across five families, and the minimal edit that re-deploys it varies with geometry: a training-free rotation matches a trained edit on the cleanest model, while distributed inversions need a trained low-rank edit, tracing a per-model correction-complexity spectrum. The cheap, self-calibrating arbiter cleanly separates grounded perception, inverted perception, and prior substitution; we argue it should be a default control for latent-knowledge and steering claims in VLMs.

Do vision models see colors the way humans do? Existing evaluations of color representations usually compare them with geometric spaces such as CIELAB or with discrete color labels. These references capture perceptual distance or category membership, but not the graded way in which people organize colors. We evaluate color grounding against a fuzzy perceptual model with 86 graded categories fitted to human survey data. The framework can be applied to any image encoder and measures three complementary properties: category boundaries, category compactness, and graded alignment beyond what color geometry alone can explain. Across eleven Vision Transformer encoders, the category-level results are broadly similar, whereas graded alignment differs substantially. Masked Autoencoders achieve the strongest beyond-geometry alignment, with confidence intervals that do not overlap those of the other encoders. A layer-wise analysis further shows that masked reconstruction preserves this structure toward the output. On natural images, MAE represents surface color globally, while language-supervised models encode color more strongly in relation to the foreground object. These results show that human-like color grounding has several distinct aspects that should not be reduced to a single score.