Explainability for Vision Foundation Models: A Survey

Foundation models create a paradox for explainability: While these powerful AI systems are increasingly used to build explainable models, their own complexity makes them inherently difficult to interpret—meaning we may be using "black boxes" to explain other systems.

Vision foundation models are particularly challenging to explain: The survey identifies that the visual domain presents unique obstacles for explainability compared to other AI applications, though the paper does not specify which architectural characteristics pose the greatest interpretability challenges.

Current evaluation methods for explainability are inconsistent: The survey reviews "common evaluation methodologies" for combined XAI and foundation model approaches, suggesting that the field lacks standardized ways to measure whether these systems are actually more understandable to users and stakeholders.

The field is rapidly evolving with significant gaps: By compiling a comprehensive corpus and categorizing existing work, the survey reveals that research bridging explainability and vision foundation models is still developing, with the authors identifying key directions for future research rather than established best practices.

Responsible AI deployment requires addressing this gap: As foundation models become more integrated into daily life, the survey underscores that we cannot rely on these systems for high-stakes applications without better understanding and communicating how they make decisions.