# Diminishing Returns in Self-Supervised Learning

arXiv: https://arxiv.org/abs/2512.03862 · HTML: https://mlsanity.com/paper/2512.03862/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 10 of its 13 references that are in the map.
- First posted on arXiv 3 Dec 2025.

## Builds on (most cited first, 10 of 10)

- [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://mlsanity.com/paper/1810.04805/) (2018)
- [ImageNet Large Scale Visual Recognition Challenge](https://mlsanity.com/paper/1409.0575/) (2014)
- [Masked Autoencoders Are Scalable Vision Learners](https://mlsanity.com/paper/2111.06377/) (2021)
- [BEiT: BERT Pre-Training of Image Transformers](https://mlsanity.com/paper/2106.08254/) (2021)
- [An Empirical Study of Training Self-Supervised Vision Transformers](https://mlsanity.com/paper/2104.02057/) (2021)
- [SimMIM: A Simple Framework for Masked Image Modeling](https://mlsanity.com/paper/2111.09886/) (2021)
- [Sentence Encoders on STILTs: Supplementary Training on Intermediate Labeled-data Tasks](https://mlsanity.com/paper/1811.01088/) (2018)
- [Are Large-scale Datasets Necessary for Self-Supervised Pre-training?](https://mlsanity.com/paper/2112.10740/) (2021)
- [Better plain ViT baselines for ImageNet-1k](https://mlsanity.com/paper/2205.01580/) (2022)
- [Rethinking Why Intermediate-Task Fine-Tuning Works](https://mlsanity.com/paper/2108.11696/) (2021)

## Abstract (arXiv)

Transformer-based architectures have become a dominant paradigm in vision and language, but their success is often attributed to large model capacity and massive training data. In this work, we examine how self-supervised pre-training, intermediate fine-tuning, and downstream fine-tuning interact in a low-capacity regime, using a 5M-parameter Vision Transformer for semantic segmentation. Across multiple data scales, we find that masked image modeling pre-training and downstream fine-tuning reliably improve performance, but with clear diminishing returns as supervision increases. In contrast, inserting an intermediate classification fine-tuning stage consistently degrades downstream performance, with the largest drops occurring precisely where pre-training is most effective. Through an analysis of patch-level representation geometry, we show that classification-based intermediate supervision actively interferes with representations learned during pre-training by collapsing spatial structure critical for dense prediction. These results indicate that, in small models, the geometry of supervision matters more than the number of training stages: misaligned intermediate objectives can negate the benefits of pre-training rather than amplify them.