Jiahao Huang and Jun Wang, University College London · VLM Interpretability

Probing the "Conscious" Mind: Selective Access and Global Workspace in Vision-Language Models

The authors introduce the Multimodal Concept Lens (MC-Lens), a multimodal Jacobian lens that maps intermediate states of a vision-language model's vision encoder and language decoder into a shared semantic space defined by the decoder's output vocabulary. Using it, they report that concept information is readable in intermediate states and selectively accessed as tasks require, and that counterfactual interventions show this information affects downstream computation. They also propose two J-space metrics for tracking cross-modal alignment during training and report shared access patterns across several VLM families and sizes.

Open the source ↗

Similar Research & Safety releases

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.