Pith. sign in

REVIEW 13 cited by

Steering CLIP's vision transformer with sparse autoencoders

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2504.08729 v1 pith:NPHMDM3J submitted 2025-04-11 cs.CV cs.AIcs.LG

Steering CLIP's vision transformer with sparse autoencoders

classification cs.CV cs.AIcs.LG
keywords visionfeaturessaesclipmodeltransformeraddressattacks
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
read the original abstract

While vision models are highly capable, their internal mechanisms remain poorly understood -- a challenge which sparse autoencoders (SAEs) have helped address in language, but which remains underexplored in vision. We address this gap by training SAEs on CLIP's vision transformer and uncover key differences between vision and language processing, including distinct sparsity patterns for SAEs trained across layers and token types. We then provide the first systematic analysis on the steerability of CLIP's vision transformer by introducing metrics to quantify how precisely SAE features can be steered to affect the model's output. We find that 10-15\% of neurons and features are steerable, with SAEs providing thousands more steerable features than the base model. Through targeted suppression of SAE features, we then demonstrate improved performance on three vision disentanglement tasks (CelebA, Waterbirds, and typographic attacks), finding optimal disentanglement in middle model layers, and achieving state-of-the-art performance on defense against typographic attacks.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 13 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. The Signs Were Always There: Training-Free Concept Detection and Steering in Raw Transformer Dimensions

    cs.LG 2026-06 conditional novelty 8.0

    Individual transformer dimensions encode concepts via signs alone, enabling training-free detection and steering without learned feature dictionaries.

  2. Can neurons speak? Semantic narration of vision at single-cell resolution

    q-bio.NC 2026-06 unverdicted novelty 7.0

    NEURRATOR bridges neural spike trains to frozen CLIP patch embeddings via a learned encoder, then uses a multimodal LM and sparse autoencoder to produce validated natural-language narrations of viewed scenes from Neur...

  3. The Signs Were Always There: Training-Free Concept Detection and Steering in Raw Transformer Dimensions

    cs.LG 2026-06 unverdicted novelty 7.0

    Sign patterns in the unrotated standard basis of transformer activations form independent binary feature registers that support training-free detection, prediction, and causal intervention across language, vision, and...

  4. From Attribution to Action: A Human-Centered Application of Activation Steering

    cs.AI 2026-04 conditional novelty 6.5

    Activation steering of SAE-attributed components lets practitioners move from correlational inspection to causal hypothesis testing on CLIP failures, with trust shifting to observed model responses (N=8 experts).

  5. Vision-Language Asymmetry in Bistable Image Captioning

    cs.CV 2026-06 unverdicted novelty 6.0

    Behavioral tests and SAE probing on 83 bistable images show simultaneous vision-tower activation of both aspects in 72% of cases, with causal steering succeeding on default-dominant but not force-balanced stimuli, loc...

  6. TEVI: Text-Conditioned Editing of Visual Representations via Sparse Autoencoders for Improved Vision-Language Alignment

    cs.CV 2026-06 unverdicted novelty 6.0

    TEVI applies sparse autoencoders and caption-conditioned masking to edit image embeddings, yielding better retrieval on MS COCO, Flickr, IIW, DOCCI, and RoCOCO benchmarks with larger gains on richer captions.

  7. Sparse Autoencoders enable Robust and Interpretable Fine-tuning of CLIP models

    cs.CV 2026-05 unverdicted novelty 6.0

    SAE-FT uses a sparse autoencoder on pre-trained CLIP visual representations to regularize fine-tuning by penalizing changes to semantically meaningful features, aiming for robust performance on ImageNet and distributi...

  8. LatentDiff: Scaling Semantic Dataset Comparison to Millions of Images

    cs.CV 2026-04 unverdicted novelty 6.0

    LatentDiff scales semantic dataset comparison to millions of images using latent spaces of vision encoders combined with sparse autoencoders and density ratio estimation, showing better accuracy and robustness than ca...

  9. From Attribution to Action: A Human-Centered Application of Activation Steering

    cs.AI 2026-04 unverdicted novelty 6.0

    Activation steering paired with attribution enables intervention-based debugging in vision models, as all 8 interviewed experts shifted to hypothesis testing, most trusted observed responses, and highlighted risks lik...

  10. Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMs

    cs.LG 2026-04 unverdicted novelty 6.0

    DACO curates a 15,000-concept dictionary from 400K image-caption pairs and uses it to initialize an SAE that enables granular, concept-specific steering of MLLM activations, raising safety scores on MM-SafetyBench and...

  11. Interpreting Video Representations with Spatio-Temporal Sparse Autoencoders

    cs.CV 2026-04 conditional novelty 6.0

    Spatio-temporal contrastive SAEs recover temporal coherence lost by hard TopK, improve action probes by +3.9% and retrieval by up to 2.8× R@1, and expose a monosemanticity metric artifact.

  12. Beyond Interpretability: When, Why, and How Sparse Autoencoders Enable Label-Free Visual Steering

    cs.CV 2025-06 unverdicted novelty 6.0

    VS2 constructs steering vectors from sparse SAE features on unlabeled in-domain activations to improve zero-shot accuracy of CLIP models by 0.93-4.12% on CIFAR-100, CUB-200, and Tiny-ImageNet while remaining forward-p...

  13. At the Edge of Understanding: Sparse Autoencoders Trace The Limits of Transformer Generalization

    cs.LG 2026-06 unverdicted novelty 4.0

    Sparse autoencoders show OOD prompts increase fallacious concept activation in transformers, offering a mechanistic measure of shift and a path to robust fine-tuning.