Vision and language foundation models, including Vision Transformers (ViTs) and Pretrained Language Models (PLMs), have achieved remarkable success across a wide range of applications, from visual recognition and question answering to natural language understanding and multimodal reasoning. However, their increasing complexity raises critical challenges in interpretability, trustworthiness, and evaluation, particularly in understanding model behavior at a conceptual level across modalities.In this dissertation, we develop a unified framework for conceptual explanations and evaluation of foundation models, bridging the gap between model representations and human-understandable concepts. Specifically, we make three major contributions:First, we propose ProbAbilistic Concept Explainers (PACE), a variational Bayesian framework for post-hoc conceptual explanation of Vision Foundation Models. PACE models the distributions of patch embeddings using a hierarchical Gaussian mixture model, producing multi-level conceptual explanations at dataset, image, and patch levels. We formally define five desiderata for trustworthy explanations—faithfulness, stability, sparsity, multi-level structure, and parsimony—and demonstrate that PACE satisfies all of them, consistently outperforming state-of-the-art baselines (ICML 2024).Second, we extend this probabilistic explanation framework to VAriational Language Concepts (VALC), addressing conceptual interpretation for Foundation Language Models. VALC treats contextual word embeddings and their associated attention weights as observations within a hierarchical Bayesian generative model, enabling concept-level interpretations at dataset, document, and word levels. Our theoretical analysis shows that VALC’s optimization is equivalent to finding optimal conceptual interpretations under the proposed desiderata (EMNLP 2024).Third, we introduce the MultiModal Needle-in-a-Haystack (MMNeedle) benchmark, which systematically evaluates the long-context retrieval capabilities of multimodal large language models. MMNeedle employs image stitching to extend visual context length and uses automated annotation protocols to generate needle-haystack pairs. Evaluation of state-of-the-art multimodal large language models reveals significant performance limitations in long-context settings, substantial frontier/open-source performance gaps, and hallucination behaviors under negative sampling (NAACL 2025).Overall, this dissertation advances the understanding of foundation models through conceptual representations and rigorous evaluation, contributing toward the development of trustworthy multimodal AI systems with clear implications for model alignment, safety, and robustness.
Hengyi Wang (Thu,) studied this question.