CVPR 20260 citations

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

Francesco Gentile, Nicola Dall'Asen, Francesco Tonini, Massimiliano Mancini, Lorenzo Vaquero, Elisa Ricci

Abstract

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.

BibTeX
@inproceedings{cvpr2026_fromweightstocon,
  title = {From Weights to Concepts: Data-Free Interpretability of CLIP via Singular Vector Decomposition},
  author = {Francesco Gentile and Nicola Dall'Asen and Francesco Tonini and Massimiliano Mancini and Lorenzo Vaquero and Elisa Ricci},
  booktitle = {CVPR 2026},
  year = {2026}
}
From Weights to Concepts: Data-Free Interpretability of CLIP via Singular Vector Decomposition · CVPR 2026