MC-Lens Probes Concept Accessibility in Vision-Language Models
November 20, 2026
The Multimodal Concept Lens uses a multimodal Jacobian to map intermediate VLM states to a shared semantic space. This method reveals how vision and language encoders access and integrate concepts within a global workspace-style architecture.
HOW THIS AFFECTS YOU
●
researcherYou can use this lens to see which specific concepts are accessible in a model's intermediate layers.