# Mechanisms of Non-Monotonic Scaling in Vision Transformers

arXiv: https://arxiv.org/abs/2511.21635 · HTML: https://mlsanity.com/paper/2511.21635/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 16 of its 18 references that are in the map.
- First posted on arXiv 26 Nov 2025.

## Builds on (most cited first, 16 of 16)

- [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://mlsanity.com/paper/2010.11929/) (2020)
- [Scaling Laws for Neural Language Models](https://mlsanity.com/paper/2001.08361/) (2020)
- [Training data-efficient image transformers & distillation through attention](https://mlsanity.com/paper/2012.12877/) (2020)
- [DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification](https://mlsanity.com/paper/2106.02034/) (2021)
- [Prevalence of Neural Collapse during the terminal phase of deep learning training](https://mlsanity.com/paper/2008.08186/) (2020)
- [Scaling Vision Transformers to 22 Billion Parameters](https://mlsanity.com/paper/2302.05442/) (2023)
- [Attention is Not All You Need: Pure Attention Loses Rank Doubly Exponentially with Depth](https://mlsanity.com/paper/2103.03404/) (2021)
- [Conditional Positional Encodings for Vision Transformers](https://mlsanity.com/paper/2102.10882/) (2021)
- [Deep ViT Features as Dense Visual Descriptors](https://mlsanity.com/paper/2112.05814/) (2021)
- [DeepViT: Towards Deeper Vision Transformer](https://mlsanity.com/paper/2103.11886/) (2021)
- [Rethinking and Improving Relative Position Encoding for Vision Transformer](https://mlsanity.com/paper/2107.14222/) (2021)
- [The Bottom-up Evolution of Representations in the Transformer: A Study with Machine Translation and Language Modeling Objectives](https://mlsanity.com/paper/1909.01380/) (2019)
- [Anti-Oversmoothing in Deep Vision Transformers via the Fourier Domain Analysis: From Theory to Practice](https://mlsanity.com/paper/2203.05962/) (2022)
- [Getting ViT in Shape: Scaling Laws for Compute-Optimal Model Design](https://mlsanity.com/paper/2305.13035/) (2023)
- [Dynamic Tuning Towards Parameter and Inference Efficiency for ViT Adaptation](https://mlsanity.com/paper/2403.11808/) (2024)
- [Guiding Neural Collapse: Optimising Towards the Nearest Simplex Equiangular Tight Frame](https://mlsanity.com/paper/2411.01248/) (2024)

## Abstract (arXiv)

Deeper Vision Transformers often perform worse than shallower ones, which challenges common scaling assumptions. Through a systematic empirical analysis of ViT-S, ViT-B, and ViT-L on ImageNet, we identify a consistent three-phase Cliff-Plateau-Climb pattern that governs how representations evolve with depth. We observe that better performance is associated with progressive marginalization of the [CLS] token, originally designed as a global aggregation hub, in favor of distributed consensus among patch tokens. We quantify patterns of information mixing with an Information Scrambling Index, and show that in ViT-L the information-task tradeoff emerges roughly 10 layers later than in ViT-B, and that these additional layers correlate with increased information diffusion rather than improved task performance. Taken together, these results suggest that transformer architectures in this regime may benefit more from carefully calibrated depth that executes clean phase transitions than from simply increasing parameter count. The Information Scrambling Index provides a useful diagnostic for existing models and suggests a potential design target for future architectures. All code is available at: https://github.com/AnanthaPadmanaban-KrishnaKumar/Cliff-Plateau-Climb.