Do vision models know that three is three?
AuthorsAnkit Sonthalia, Arnas Uselis, Seong Joon Oh
Vision models often learn a concept change clearly within one visual setting, but may represent that same change differently in another.
Key results
Models spanning architectures and training objectives.
Included numerosity, age, object identity, and visual attributes.
Local direction consistency exceeded the random baseline.
Global direction consistency exceeded the random baseline.
Upper end of the association between local and global linearity across concepts.
What the paper found
Do vision models represent “three” as the same transformation in handwritten digits, finger counts, and other visual settings? The study tests the linear representation hypothesis across 46 vision encoders and 10 concept families, including numerosity, age, object identity, and visual attributes. It estimates concept directions from matched image pairs and checks whether their embedding changes align within and between visual regimes, against random baselines. Local linearity is nearly universal, but those directions often fail to transfer across regimes—a pattern the paper calls “fragmented linearity.” Across all tests, 99.1% of local comparisons were significant, versus 52.2% of cross-regime comparisons. The models include OpenAI CLIP variants, SigLIP, and DINOv3; CLIP examples show stronger cross-regime alignment than supervised models. Local and global linearity were positively associated across every concept, with Kendall’s τ ranging from 0.38 to 0.89. Where directions align globally, they can also transfer as linear probes, suggesting that some vision encoders capture regime-independent concepts rather than learning separate visual rules for each setting.
Original 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.
Read the original paperMore in Neural Networks
Browse all 28 papers →Curriculum unlearning in humans and neural networks
Jirko Rubruck, Alice Zhang, Andrew Saxe, Christopher Summerfield
Simple-to-complex curricula may work by helping humans and neural networks unlearn overly simple rules and build richer representations.
Meta Models: Models that Model Learning
Peihang Li, Ziming Liu
Instead of running a model-training experiment first, the authors use prior runs to predict how the next one will learn.
Continuous Memory Machines
Ciaran Regan, Kai Arulkumaran, Luke Darlow, Stefania Druga, Sebastian Risi, Llion Jones
Continuous Memory Machines give recurrent networks separate short- and long-term memories that can read from and reorganize one another.