Explaining, Verifying, and Aligning Semantic Hierarchies in Vision-Language Model Embeddings

A team of researchers including Gesina Schwalbe, Mert Keser, and nine others has introduced a novel framework for dissecting the 'mental models' of popular Vision-Language Models (VLMs) like CLIP. These models create a shared embedding space for images and text, but their internal semantic organization—how they categorize concepts like 'dog' under 'animal'—has been a black box. The new method performs post-hoc analysis by first extracting a binary hierarchy from class centroids via agglomerative clustering and naming the nodes using a concept bank. It then quantifies how plausible this AI-generated tree is by comparing it to human ontologies using specialized tree- and edge-level consistency measures.

The framework's utility is demonstrated through explainable hierarchical inference, which includes uncertainty-aware early stopping (UAES). Crucially, the researchers also propose an alignment method that can post-hoc tweak a VLM's embedding space using a lightweight transformation, guided by a target hierarchy generated via techniques like UMAP. Their sweeping analysis across 13 VLMs and 4 image datasets uncovered a fundamental, persistent trade-off: the image encoder pathways are more discriminative and yield higher zero-shot accuracy, while the text encoder pathways induce semantic hierarchies that are more ontologically plausible and better match human taxonomies.

This finding has significant implications for AI development, suggesting that optimizing purely for task accuracy might come at the cost of building models that 'think' in ways understandable to humans. The work provides practical tools for developers to diagnose and improve the semantic alignment of their models, moving beyond benchmark scores to evaluate the conceptual soundness of AI systems. It opens a new avenue for creating more interpretable and trustworthy multimodal AI.