# Large Vision Models Can Solve Mental Rotation Problems

arXiv: https://arxiv.org/abs/2509.15271 · HTML: https://mlsanity.com/paper/2509.15271/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- Placed in Spatial and embodied reasoning, one of 17 directions in LLM reasoning.
- Builds on 21 of its 25 references that are in the map.
- First posted on arXiv 18 Sep 2025.

## Placed in

- LLM reasoning / [Spatial and embodied reasoning](https://mlsanity.com/lr/spatial-and-embodied-reasoning/)

## Builds on (most cited first, 21 of 21)

- [Learning Transferable Visual Models From Natural Language Supervision](https://mlsanity.com/paper/2103.00020/) (2021)
- [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://mlsanity.com/paper/2010.11929/) (2020)
- [DINOv2: Learning Robust Visual Features without Supervision](https://mlsanity.com/paper/2304.07193/) (2023)
- [Masked Autoencoders Are Scalable Vision Learners](https://mlsanity.com/paper/2111.06377/) (2021)
- [Spatial Transformer Networks](https://mlsanity.com/paper/1506.02025/) (2015)
- [DINOv3](https://mlsanity.com/paper/2508.10104/) (2025)
- [Reproducible scaling laws for contrastive language-image learning](https://mlsanity.com/paper/2212.07143/) (2022)
- [Group Equivariant Convolutional Networks](https://mlsanity.com/paper/1602.07576/) (2016)
- [How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers](https://mlsanity.com/paper/2106.10270/) (2021)
- [Conditional Positional Encodings for Vision Transformers](https://mlsanity.com/paper/2102.10882/) (2021)
- [Layer by Layer: Uncovering Hidden Representations in Language Models](https://mlsanity.com/paper/2502.02013/) (2025)
- [Intrinsic dimension of data representations in deep neural networks](https://mlsanity.com/paper/1905.12784/) (2019)
- [Learning SO(3) Equivariant Representations with Spherical CNNs](https://mlsanity.com/paper/1711.06721/) (2017)
- [Rotation equivariant vector field networks](https://mlsanity.com/paper/1612.09346/) (2016)
- [From Colors to Classes: Emergence of Concepts in Vision Transformers](https://mlsanity.com/paper/2503.24071/) (2025)
- [You Don't Need Domain-Specific Data Augmentations When Scaling Self-Supervised Learning](https://mlsanity.com/paper/2406.09294/) (2024)
- [Beyond Scalars: Concept-Based Alignment Analysis in Vision Transformers](https://mlsanity.com/paper/2412.06639/) (2024)
- [The Geometry of Self-supervised Learning Models and its Impact on Transfer Learning](https://mlsanity.com/paper/2209.08622/) (2022)
- [On convex decision regions in deep network representations](https://mlsanity.com/paper/2305.17154/) (2023)
- [Humans and deep networks largely agree on which kinds of variation make object recognition harder](https://mlsanity.com/paper/1604.06486/) (2016)
- [Connecting Concept Convexity and Human-Machine Alignment in Deep Neural Networks](https://mlsanity.com/paper/2409.06362/) (2024)

## Built on by (most cited first, 4 of 4)

- [SeeSE3: Emergence of 3D Space in Vision Features](https://mlsanity.com/paper/2607.14228/) (2026)
- [Semantic Richness or Geometric Reasoning? The Fragility of VLM's Visual Invariance](https://mlsanity.com/paper/2604.01848/) (2026)
- [LEVANTE-bench: Multi-Scale Comparison of VLMs to Children Using Cognitive Tasks (or, "Is Your VLM Smarter Than a 5th Grader?")](https://mlsanity.com/paper/2606.05497/) (2026)
- [Vision Transformers Learn Word Meaning from Images Alone](https://mlsanity.com/paper/2610.13252/) (2026)

## Abstract (arXiv)

Mental rotation is a key test of spatial reasoning in humans and has been central to understanding how perception supports cognition. Despite the success of modern vision transformers, it is still unclear how well these models develop similar abilities. In this work, we present a systematic evaluation of ViT, CLIP, DINOv2, and DINOv3 across a range of mental-rotation tasks, from simple block structures similar to those used by Shepard and Metzler to study human cognition, to more complex block figures, three types of text, and photo-realistic objects. By probing model representations layer by layer, we examine where and how these networks succeed. We find that i) self-supervised ViTs capture geometric structure better than supervised ViTs; ii) intermediate layers perform better than final layers; iii) task difficulty increases with rotation complexity and occlusion, mirroring human reaction times and suggesting similar constraints in embedding space representations.