| title | Sparse Autoencoders enable Robust and Interpretable Fine-tuning of CLIP models |
|---|---|
| authors | Fabian Morelli, Arnas Uselis, Ankit Sonthalia, Seong Joon Oh |
| date | 2026-05-15 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2605.15961 |
| upvotes | 6 (Hugging Face Daily Papers, 2026-11-01) |
| links | abs pdf hf code |
| abstract | Large-scale pre-trained vision-language models like CLIP demonstrate remarkable zero-shot performance across diverse tasks. However, fine-tuning these models to improve downstream performance often degrades robustness against distribution shifts. Recent approaches have attempted to mitigate this trade-off, but often rely on computationally expensive text-guidance. We propose a novel method for robust fine-tuning, SAE-FT, which operates only on the model's visual representations. SAE-FT regularizes changes to these representations by penalizing the addition and removal of semantically meaningful features identified by a Sparse Autoencoder trained on the pre-trained model. This constraint prevents catastrophic forgetting and makes the fine-tuning process interpretable, enabling direct analysis of semantic changes. SAE-FT is both mechanistically transparent and computationally efficient, matching or exceeding state-of-the-art performance on ImageNet and its associated distribution shift benchmarks. Code is publicly available at: https://github.com/Fabian-Mor/sae-ft. |
cite
@misc{morelli2026sparse,
title={Sparse Autoencoders enable Robust and Interpretable Fine-tuning of CLIP models},
author={Fabian Morelli and Arnas Uselis and Ankit Sonthalia and Seong Joon Oh},
year={2026},
eprint={2605.15961},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2605.15961}
}