Do vision models know that three is three?

Ankit Sonthalia, Arnas Uselis, Seong Joon OharXiv 2026 · arXiv

TL;DR

Vision models represent abstract concepts (like the number "3") as linear directions in their embedding space, but these directions are often "fragmented"—consistent within a specific visual style (e.g., handwritten digits) but different in another (e.g., hand gestures). As models scale and use more diverse training objectives (like CLIP), these fragmented directions converge into a single, globally consistent "concept vector," supporting the idea that AI is converging toward a shared, abstract model of reality.

What Problem Does This Paper Address?

The Linear Representation Hypothesis (LRH) suggests that high-level concepts are encoded as simple linear directions in a model's internal embedding space. While well-documented in Large Language Models (LLMs), vision presents a unique challenge: the same concept can look radically different depending on the context. For example, the concept of "three" can manifest as the digit "3," three fingers, or three dots. These "regimes" share almost nothing in common in pixel space.

The authors ask: Do vision models recognize the underlying abstraction? Does the vector for "adding one" to a set of dots look the same as the vector for "adding one" to a hand gesture? Understanding this is crucial for model interpretability and for building multimodal systems that truly "understand" the world rather than just memorizing visual patterns.

Methods

The researchers evaluated 46 different vision encoders (including CLIP, DINO, MAE, and supervised ResNets/ViTs) across 10 concept families. To measure linearity, they introduced several key frameworks:

  • Regimes: They partitioned data into visually distinct groups (e.g., for the concept of "numerosity," regimes included MNIST digits, hand fingers, and rendered spheres).
  • Approximate Counterfactuals: Unlike text, you can't easily swap one concept in an image while keeping everything else identical. The authors used linear sum assignment to find pairs of images within a regime that were as close as possible in embedding space but differed in the target concept (e.g., a "2" and a "3" with similar stroke thickness).
  • Direction Consistency (DC): A metric that measures how well these counterfactual differences align with a single common direction. They distinguished between Local DC (consistency within one regime) and Cross-Regime DC (whether a direction from one regime works in another).

Datasets

The study utilized a diverse array of benchmarks to test 10 concept families:

  • Numerosity: MNIST (digits), Fingers, CLEVR (objects), and a synthetic dot-counting dataset.
  • Age: UTKFace and 11k-Hands.
  • Object Identity: The PACS dataset (Photos, Art, Cartoons, Sketches).
  • MIT-States Attributes: 7 binary attributes including Clean/Dirty, Dry/Wet, and Raw/Cooked.

Key Insights and Ideas

  • Fragmented Linearity: This is the paper's central discovery. Many models exhibit strong "local" linearity (they have a clear vector for "3" within MNIST) but "fragmented" global linearity (the MNIST "3" vector doesn't work for the Fingers regime). As shown in Figure 1, a model can be locally linear but globally "lost."
Figure 1: Vision embeddings exhibit fragmented linearity. (a) Three possibilities for linearity: none, fragmented, or global. (b) Empirically, we often observe fragmented linearity.
  • The Convergence Trend: There is a strong positive correlation ($\tau \in [0.38, 0.89]$) between local and global linearity. As models get better at representing a concept consistently in one context, they naturally begin to align those representations across all contexts.
  • Platonic Convergence: The results support the "Platonic Representation Hypothesis"—the idea that as models become more powerful, their internal representations of the world converge toward a shared, objective ideal, regardless of their specific training data or architecture.
  • Concrete vs. Abstract: Concepts with consistent visual cues (like "Age" via wrinkles or "Object Identity" via shape) show much higher global linearity than abstract concepts like "Numerosity" or "Cleanliness," where the visual manifestation varies wildly.

Key Results

  • Local linearity is nearly universal: 99.1% of all (model, concept, regime) triplets showed statistically significant local linearity.
  • Global linearity is the differentiator: Only 52.2% of tests passed for global linearity. Supervised models were the most fragmented, while CLIP-style models were "linear almost everywhere."
  • CLIP vs. Supervised: As seen in Figure 2, a supervised ViT might have a fragmentation score of 0.67 for numerosity, while CLIP achieves a score of 0 (perfect global consistency).
Figure 2: Within-regime (diagonal) and cross-regime (off-diagonal) direction consistencies. Some values are indistinguishable from baseline (according to Welch’s t-test with Bonferroni-corrected p-values) and are excluded from display. Even supervised models (leftmost panel) are locally linear, while only selectively globally linear. In contrast, CLIP models are linear almost everywhere (most off-diagonal entries significant).
  • Linear Probing: In globally linear models, a concept direction discovered in one regime (e.g., "clean vs. dirty" in kitchens) can act as a highly accurate linear probe in a completely different regime (e.g., carpets), as shown in Figure 4.
Figure 4: Probe accuracy vs. global linearity. Higher global linearity directly translates to better cross-regime concept classification.

Limitations and Future Work

The authors acknowledge that their counterfactual pairs are approximate, relying on embedding distance which might not perfectly isolate a single concept. Furthermore, the study is limited to 10 concept families; more complex or nuanced concepts might behave differently.

Future work could explore whether these linear directions can be used to "steer" vision models (e.g., making an image "cleaner" in representation space) and whether the emergence of global linearity is a prerequisite for the advanced reasoning seen in the latest multimodal models.

Summary updated

Abstract

Do vision models encode the transition $2{\to}3$ 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 $τ\in [0.38, 0.89]$ across all concepts), and where global linearity does emerge, concept directions can transfer as linear probes across regimes.
BibTeX
@misc{sonthalia2026do,
  author = {Ankit Sonthalia and Arnas Uselis and Seong Joon Oh},
  title = {{Do vision models know that three is three?}},
  year = {2026},
  eprint = {2610.35857},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.35857},
}
Figure