cs.CVNov 14, 2026

Beyond the Training Cameras: Coverage-Controlled Probes of Viewing Direction in Frozen Vision Models

Authors: Peize Ding

Organizations: Columbia University

Abstract

Vision models are trained on 2D images of a 3D world. Do their hidden features encode the direction from which the camera views a scene, and if so, in what form? We study the relative viewing direction αα of two objects in rendered scenes where both object centres stay at fixed pixels while the camera moves, so that perspective, shading and shadows change but the projected layout does not. In eight pretrained models, from diffusion models to self-supervised and image-text encoders, αα is encoded in a strikingly simple form: a single two-dimensional plane in the hidden space, shared across scenes, in which the difference between the two objects' features traces (cos⁡α,sin⁡α)(\cosα, \sinα). The plane keeps its orientation when lighting, colours, floors and object shapes change. This encoding is more than a summary of a few visible measurements: readouts fitted on one set of camera paths stay accurate at camera positions outside their range in seven of the eight models, whereas decoders of the measured object and shadow geometry, equally accurate within that range, fail there. In the diffusion transformer PixArt-αα, the plane forms abruptly at a single block and carries direction information to later blocks, yet editing it changes the objects rather than the generated viewpoint. We also examine how this encoding carries over to scanned real objects.

Figures & tables

Explore similar work

Mar 6, 2026

Do Foundation Models Know Geometry? Probing Frozen Features for Continuous Physical Measurement

Vision-language models encode continuous geometry that their text pathway fails to express: a 6,000-parameter linear probe extracts hand joint angles at 6.1 degrees MAE from frozen features, while the best text output achieves only 20.0 degrees -- a 3.3x bottleneck. LoRA fine-tuning (r=16, 2,000 images) narrows this gap to 6.5 degrees, providing evidence for a pathway-training deficit rather than a representational one. Training objective determines accuracy more than architecture: five encoders spanning self-supervised, contrastive, and hybrid paradigms converge to statistically equivalent accuracy (R^2 approximately 0.55, TOST-equivalent at delta=0.03) despite sharing as little as CKA=0.41 representational similarity -- functional convergence without representational convergence. Autoregressive generation damages geometric fidelity, but the damage originates in the generation process, not in language alignment: Qwen2.5-VL's LLM layers actually improve probe accuracy over its raw vision encoder. Layer-wise analysis reveals a universal mid-network accuracy peak across all architectures, with attention heads in layers 18-22 carrying disproportionate geometric signal. These findings enable a single frozen backbone to function as a multi-task geometric sensor through lightweight probes, without fine-tuning or text generation.
Oct 5, 2026

Readout Blindness: VLM Scores Miss the Spatial Direction Their Frozen Encoders Retain

CLIP-like vision-language models remain a cornerstone of multimodal systems, yet their scores stay near chance on directed spatial relations, such as whether one object is left of another. We call this failure readout blindness and analyze, theoretically and empirically, why deployed scores miss the direction: when scoring rules treat the subject and object symmetrically, direction cancels regardless of encoder training. Guided by this analysis, we introduce Antisymmetric Displacement Readout (ADR), which aligns caption words with image patches in the frozen features and scores each relation by the signed displacement between matched object centroids. Notably, ADR succeeds without additional training or learned parameters, thereby demonstrating that directional information remains in the frozen encoder. However, text and world priors can inflate accuracy, so we further introduce prior deflation, which measures the benefit of the image-text pairing as the grounded gain over a null that pairs each item with an unrelated image. Extensive experiments across encoder families show that ADR substantially improves over deployed scores, which remain near chance on most direction-balanced sets even for fine-tuned encoders. Compared with more complex readouts, ADR outperforms the evaluated MLLM likelihood readouts and is competitive with their chat inference at a small fraction of the computation. These results support our claim that directional information can be recovered from frozen features by an appropriate readout. Our implementation and evaluation kit will be publicly available.
Oct 26, 2026

Do vision models know that three is three?

Do vision models encode the transition 2→32{\to}3 the same way whether it appears in handwritten digits or in hand gestures, and do they "understand" that clean→{\to}dirty is the same transformation regardless of whether it is applied to a kitchen or a carpet? The linear representation hypothesis (LRH) answers this for language models: concept changes correspond to directions in (language) embedding space. In vision, the same concept can manifest in visually unrelated "regimes" that rarely co-occur: images of digits and hands do not share much in common in pixel space. To understand whether vision models still encode the underlying abstract concepts, we study 46 vision encoders across architectures, training objectives, and scales, on 10 concept families including numerosity, age, object identity, and MIT-States attributes. Within each regime, we find statistically significant concept directions near-universally across our tested models. However, two distinct regimes do not always share the same concept direction. Several models learn multiple linear structures for the same concept transition, split across regimes. We call this phenomenon "fragmented linearity": strong local linearity coexisting with near-absent global linearity. Fragmentation is, however, not fixed: models with stronger local linearity progressively escape it (local-global linearity correlated with τ∈[0.38,0.89]τ\in [0.38, 0.89] across all concepts), and where global linearity does emerge, concept directions can transfer as linear probes across regimes.