From Weights to Concepts: Data-Free Interpretability of CLIP via Singular Vector Decomposition

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Gentile, Francesco, Dall'Asen, Nicola, Tonini, Francesco, Mancini, Massimiliano, Vaquero, Lorenzo, Ricci, Elisa
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866910073345802240
author Gentile, Francesco
Dall'Asen, Nicola
Tonini, Francesco
Mancini, Massimiliano
Vaquero, Lorenzo
Ricci, Elisa
author_facet Gentile, Francesco
Dall'Asen, Nicola
Tonini, Francesco
Mancini, Massimiliano
Vaquero, Lorenzo
Ricci, Elisa
contents As vision-language models are deployed at scale, understanding their internal mechanisms becomes increasingly critical. Existing interpretability methods predominantly rely on activations, making them dataset-dependent, vulnerable to data bias, and often restricted to coarse head-level explanations. We introduce SITH (Semantic Inspection of Transformer Heads), a fully data-free, training-free framework that directly analyzes CLIP's vision transformer in weight space. For each attention head, we decompose its value-output matrix into singular vectors and interpret each one via COMP (Coherent Orthogonal Matching Pursuit), a new algorithm that explains them as sparse, semantically coherent combinations of human-interpretable concepts. We show that SITH yields coherent, faithful intra-head explanations, validated through reconstruction fidelity and interpretability experiments. This allows us to use SITH for precise, interpretable weight-space model edits that amplify or suppress specific concepts, improving downstream performance without retraining. Furthermore, we use SITH to study model adaptation, showing how fine-tuning primarily reweights a stable semantic basis rather than learning entirely new features.
format Preprint
id arxiv_https___arxiv_org_abs_2603_24653
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle From Weights to Concepts: Data-Free Interpretability of CLIP via Singular Vector Decomposition
Gentile, Francesco
Dall'Asen, Nicola
Tonini, Francesco
Mancini, Massimiliano
Vaquero, Lorenzo
Ricci, Elisa
Computer Vision and Pattern Recognition
As vision-language models are deployed at scale, understanding their internal mechanisms becomes increasingly critical. Existing interpretability methods predominantly rely on activations, making them dataset-dependent, vulnerable to data bias, and often restricted to coarse head-level explanations. We introduce SITH (Semantic Inspection of Transformer Heads), a fully data-free, training-free framework that directly analyzes CLIP's vision transformer in weight space. For each attention head, we decompose its value-output matrix into singular vectors and interpret each one via COMP (Coherent Orthogonal Matching Pursuit), a new algorithm that explains them as sparse, semantically coherent combinations of human-interpretable concepts. We show that SITH yields coherent, faithful intra-head explanations, validated through reconstruction fidelity and interpretability experiments. This allows us to use SITH for precise, interpretable weight-space model edits that amplify or suppress specific concepts, improving downstream performance without retraining. Furthermore, we use SITH to study model adaptation, showing how fine-tuning primarily reweights a stable semantic basis rather than learning entirely new features.
title From Weights to Concepts: Data-Free Interpretability of CLIP via Singular Vector Decomposition
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2603.24653