Future Direction: Multimodal Mechanistic Interpretability

Background

Current mechanistic interpretability and white-box monitoring are predominantly text-based. Multimodal systems introduce new failure modes where visual and textual signals may conflict or be misrouted.

Content

Extend the “knowing vs. doing” framework to vision-language models: probe when and why multimodal representations fail to route into downstream decisions.

Core Contribution

Build white-box monitoring and intervention tools for multimodal RAG and agentic systems, ensuring that visual evidence governs generation as reliably as textual evidence.

Vision

Safe AI for high-stakes deployment must be inspectable, intervenable, and accountable across all modalities.