# SpIn-ViT: Designing a Sparsity-Induced Vision Transformer That Is Mechanistically Interpretable

arXiv: https://arxiv.org/abs/2608.14922 · HTML: https://mlsanity.com/paper/2608.14922/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Interpreting vision and multimodal models, one of 13 directions in Mechanistic interpretability.
- Builds on 18 of its 36 references that are in the map.
- First posted on arXiv 14 Aug 2026.

## Placed in

- Mechanistic interpretability / [Interpreting vision and multimodal models](https://mlsanity.com/mechinterp/interpreting-vision-and-multimodal-models/)

## Builds on (most cited first, 18 of 18)

- [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)
- [End-to-End Object Detection with Transformers](https://mlsanity.com/paper/2005.12872/) (2020)
- [Training data-efficient image transformers & distillation through attention](https://mlsanity.com/paper/2012.12877/) (2020)
- [Deformable DETR: Deformable Transformers for End-to-End Object Detection](https://mlsanity.com/paper/2010.04159/) (2020)
- [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://mlsanity.com/paper/2105.15203/) (2021)
- [Describing Textures in the Wild](https://mlsanity.com/paper/1311.3618/) (2013)
- [Fine-Grained Visual Classification of Aircraft](https://mlsanity.com/paper/1306.5151/) (2013)
- [EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification](https://mlsanity.com/paper/1709.00029/) (2017)
- [Refusal in Language Models Is Mediated by a Single Direction](https://mlsanity.com/paper/2406.11717/) (2024)
- [Segmenter: Transformer for Semantic Segmentation](https://mlsanity.com/paper/2105.05633/) (2021)
- [RISE: Randomized Input Sampling for Explanation of Black-box Models](https://mlsanity.com/paper/1806.07421/) (2018)
- [Sparse autoencoders reveal selective remapping of visual concepts during adaptation](https://mlsanity.com/paper/2412.05276/) (2024)
- [Aligning Large Language Models with Representation Editing: A Control Perspective](https://mlsanity.com/paper/2406.05954/) (2024)
- [Spectral Editing of Activations for Large Language Model Alignment](https://mlsanity.com/paper/2405.09719/) (2024)
- [ERIC: Extracting Relations Inferred from Convolutions](https://mlsanity.com/paper/2010.09452/) (2020)
- [Symbolic Rule Extraction from Attention-Guided Sparse Representations in Vision Transformers](https://mlsanity.com/paper/2505.06745/) (2025)
- [Improving Interpretability and Accuracy in Neuro-Symbolic Rule Extraction Using Class-Specific Sparse Filters](https://mlsanity.com/paper/2501.16677/) (2025)

## Abstract (arXiv)

Mechanistic interpretability has recently expanded to Vision Transformers (ViTs), with Sparse Autoencoders (SAEs) increasingly used as post-hoc tools to decompose internal representations into sparse and more interpretable features. However, because post-hoc SAEs are trained on frozen representations after the ViT has already been optimized, their latent features are not directly aligned with the downstream classification objective. We introduce SpIn-ViT, a framework that jointly trains a pretrained ViT and a modified SAE end-to-end, directly aligning sparse patch-level representations with image classification. SpIn-ViT learns semantically coherent neuron activations that localize meaningful image regions while maintaining competitive predictive performance. We evaluate SpIn-ViT across nine image-classification benchmarks using classification accuracy, quantitative interpretability metrics, AI-based and Human evaluations. Compared with the previous state-of-the-art post-hoc SAE method, SpIn-ViT achieves 8.84% higher average classification accuracy, an AI-based interpretability score nearly four times as high, and a human-evaluation score more than twice as high. We further extract interpretable rule-sets using the SAE neurons to create neurosymbolic models which achieve 5.97% higher average classification accuracy while requiring a 58.8% smaller rule-set than the neurosymbolic models created from the SOTA post-hoc SAE method.