Concept Regions Matter: Benchmarking CLIP with a New Cluster-Importance Approach

2025-11-17

AI Summary

Vision-language models (VLMs) like CLIP suffer from background-based spurious correlations in zero-shot recognition, yet existing evaluation protocols cannot disentangle performance degradation caused by background bias from confounding factors such as viewpoint/scale variation or fine-grained confusion. Method: We propose Cluster-based Concept Importance (CCI), an interpretable attribution method that leverages CLIP’s patch embeddings for semantic clustering, integrates masked-response analysis, and incorporates GroundedSAM-based foreground-background segmentation to quantify concept importance. Contribution/Results: Building upon CCI, we introduce COVAR—the first controlled-variable benchmark that systematically isolates multiple interference sources. Experiments show CCI improves deletion-AUC on MS COCO by over 2×, setting a new state-of-the-art in attribution faithfulness. COVAR’s evaluation of 18 CLIP variants exposes critical limitations of prevailing debiasing benchmarks, providing both diagnostic tools and rigorous evaluation standards for developing robust VLMs.

Technology Category

• Computer Vision: Interpretability, Explainability, and Transparency
• Machine Learning: Large Multimodal Models (LMMs)
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Contrastive vision-language models (VLMs) such as CLIP achieve strong zero-shot recognition yet remain vulnerable to spurious correlations, particularly background over-reliance. We introduce Cluster-based Concept Importance (CCI), a novel interpretability method that uses CLIP's own patch embeddings to group spatial patches into semantically coherent clusters, mask them, and evaluate relative changes in model predictions. CCI sets a new state of the art on faithfulness benchmarks, surpassing prior methods by large margins; for example, it yields more than a twofold improvement on the deletion-AUC metric for MS COCO retrieval. We further propose that CCI, when combined with GroundedSAM, automatically categorizes predictions as foreground- or background-driven, providing a crucial diagnostic ability. Existing benchmarks such as CounterAnimals, however, rely solely on accuracy and implicitly attribute all performance degradation to background correlations. Our analysis shows this assumption to be incomplete, since many errors arise from viewpoint variation, scale shifts, and fine-grained object confusions. To disentangle these effects, we introduce COVAR, a benchmark that systematically varies object foregrounds and backgrounds. Leveraging CCI with COVAR, we present a comprehensive evaluation of eighteen CLIP variants, offering methodological advances and empirical evidence that chart a path toward more robust VLMs.

Problem

• Evaluating CLIP's vulnerability to spurious background correlations
• Proposing Cluster-based Concept Importance for interpretable spatial analysis
• Introducing COVAR benchmark to disentangle foreground and background effects

Innovation

• Cluster-based Concept Importance method for interpretability
• Combining CCI with GroundedSAM for diagnostic categorization
• Introducing COVAR benchmark for systematic foreground-background variation