# Interpreting vision transformers via residual replacement model

arXiv: https://arxiv.org/abs/2509.17401 · HTML: https://mlsanity.com/paper/2509.17401/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 22 Sep 2025.

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

- [Interpretable Debiasing of Vision-Language Models for Social Fairness](https://mlsanity.com/paper/2602.24014/) (2026)
- [Sparse Autoencoders are Topic Models](https://mlsanity.com/paper/2511.16309/) (2025)
- [Rethinking Prototype-based Similarity Learning for Few-Shot Object Detection](https://mlsanity.com/paper/2606.23069/) (2026)
- [FailSAE: Towards Interpretable Failure Prediction for Vision-Language Models via Sparse Autoencoders](https://mlsanity.com/paper/2609.04276/) (2026)

## Abstract (arXiv)

How do vision transformers (ViTs) represent and process the world? This paper addresses this long-standing question through the first systematic analysis of 6.6K features across all layers, extracted via sparse autoencoders, and by introducing the residual replacement model, which replaces ViT computations with interpretable features in the residual stream. Our analysis reveals not only a feature evolution from low-level patterns to high-level semantics, but also how ViTs encode curves and spatial positions through specialized feature types. The residual replacement model scalably produces a faithful yet parsimonious circuit for human-scale interpretability by significantly simplifying the original computations. As a result, this framework enables intuitive understanding of ViT mechanisms. Finally, we demonstrate the utility of our framework in debiasing spurious correlations.