Pith. sign in

REVIEW 17 cited by

Interpreting CLIP's Image Representation via Text-Based Decomposition

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 2310.05916 v4 pith:ZF243VS5 submitted 2023-10-09 cs.CV cs.AI

classification cs.CVcs.AI
keywords imagecliprepresentationheadsinterpretingattentionindividualmodel
verification ladder T0 review T1 audit T2 compute T3 formal

Signed reviews

No signed human review yet.

0 comments
read the original abstract

We investigate the CLIP image encoder by analyzing how individual model components affect the final representation. We decompose the image representation as a sum across individual image patches, model layers, and attention heads, and use CLIP's text representation to interpret the summands. Interpreting the attention heads, we characterize each head's role by automatically finding text representations that span its output space, which reveals property-specific roles for many heads (e.g. location or shape). Next, interpreting the image patches, we uncover an emergent spatial localization within CLIP. Finally, we use this understanding to remove spurious features from CLIP and to create a strong zero-shot image segmenter. Our results indicate that a scalable understanding of transformer models is attainable and can be used to repair and improve models.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 17 Pith papers

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

  1. Multimodal Model Diffing for Feature Discovery and Control

    cs.CV 2026-08 conditional novelty 7.0 of 10

    By diffing base-language and multimodal sparse autoencoder features, MMDiff isolates causally relevant features that can be ablated or steered to control spatial, OCR, and safety behaviors in multimodal LLMs.

  2. ConceptAttention: Diffusion Transformers Learn Highly Interpretable Features

    cs.CV 2025-02 conditional novelty 7.0 of 10

    ConceptAttention shows that linear projections in the output space of DiT attention layers yield sharper concept-localizing saliency maps than cross-attention maps, reaching state-of-the-art zero-shot segmentation.

  3. Prune Once: Retraining-Free Task-Agnostic Pruning for Vision-Language Models

    cs.CV 2026-08 conditional novelty 6.0 of 10

    A variance-based, retraining-free pruning framework for vision-language models that allocates per-layer sparsity and outperforms Wanda and SparseGPT at high sparsity.

  4. BrainExplore: Large-Scale Discovery of Interpretable Visual Representations in the Human Brain

    cs.CV 2025-12 conditional novelty 6.0 of 10

    A new automated pipeline decomposes fMRI activity into components and labels them with visual concepts, claiming thousands of interpretable patterns across the human visual cortex.

  5. The SuperActivator Mechanism: Transformers Concentrate Reliable Concept Signals in the Tail

    cs.LG 2025-12 conditional novelty 6.0 of 10

    Reliable concept presence in transformers is concentrated in the extreme high-activation tail of in-concept tokens; thresholding that tail improves concept detection and localization.

  6. How Visual Representations Map to Language Feature Space in Multimodal LLMs

    cs.CV 2025-06 conditional novelty 6.0 of 10

    Visual tokens in a fully frozen-backbone VLM with a linear adapter only become well-represented by the LLM's sparse autoencoder features in middle-to-late layers, converging around layer 18.

  7. Debiasing CLIP: Interpreting and Correcting Bias in Attention Heads

    cs.CV 2025-05 conditional novelty 6.0 of 10

    Using wrong/correct hard-sample head comparisons, LTC finds spurious CLIP attention heads and corrects them to raise worst-group accuracy on biased benchmarks.

  8. Understanding Design Fixation in Generative AI

    cs.HC 2025-02 conditional novelty 6.0 of 10

    Generative AI models exhibit a design fixation phenomenon that limits the diversity and originality of their design outputs, according to a small lab study and a proposed theoretical framework.

  9. Padding Tone: A Mechanistic Analysis of Padding Tokens in T2I Models

    cs.CL 2025-01 conditional novelty 6.0 of 10

    Padding tokens in text-to-image models can carry semantic information or act as diffusion-time registers, depending on training and attention architecture.

  10. Evaluating Sample Utility for Efficient Data Selection by Mimicking Model Weights

    cs.LG 2025-01 conditional novelty 6.0 of 10

    A gradient alignment score against a reference model's weights selects and reweights training samples, improving data efficiency in image classification and CLIP pretraining.

  11. Cross-Attention Head Position Patterns Can Align with Human Visual Concepts in Text-to-Image Generative Models

    cs.CV 2024-12 conditional novelty 6.0 of 10

    Head Relevance Vectors, built by counting which attention head best matches each of 34 concept words, reveal concept-specific cross-attention head patterns and enable targeted concept steering in Stable Diffusion.

  12. Understanding Multimodal LLMs: the Mechanistic Interpretability of Llava in Visual Question Answering

    cs.CL 2024-11 conditional novelty 6.0 of 10

    The paper shows LLaVA's visual QA mechanism parallels textual QA: visual embeddings encode animal and color features, attention heads extract and match them, and visual instruction tuning refines existing Vicuna heads.

  13. MEDIC-AD: Towards Medical Vision-Language Model's Clinical Intelligence

    cs.CV 2026-03 reject novelty 5.0 of 10

    MEDIC-AD adds anomaly-aware and difference tokens to a medical VLM, claiming SOTA lesion detection, temporal tracking, and visual grounding; the zero-shot claim is undermined by likely train/test overlap.

  14. Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning

    cs.CV 2025-01 reject novelty 5.0 of 10

    Neuron dissection of the infant-trained CVCL model claims to find visual concepts beyond its training vocabulary, but the evaluation leaks test data into neuron selection.

  15. Model Science: getting serious about verification, explanation and control of AI systems

    cs.AI 2025-08 conditional novelty 4.0 of 10

    Proposes 'Model Science' as a model-centric paradigm for AI with four pillars: verification, explanation, control, and interface.

  16. A Review of Multimodal Explainable Artificial Intelligence: Past, Present and Future

    cs.CV 2024-12 conditional novelty 4.0 of 10

    A historical review that organizes multimodal explainability methods into four chronological eras and three explainability types, extending coverage to generative LLMs.

  17. Explainable and Interpretable Multimodal Large Language Models: A Comprehensive Survey

    cs.CL 2024-12 conditional novelty 4.0 of 10

    A survey maps the field of MLLM explainability and interpretability into data, model, and training and inference perspectives.

Pith tools