MC-Lens probes selective access and global workspace functions inside vision-language models

Jiahao Huang · hf · 2026-11-20

Imperial College London and UCL researchers introduce MC-Lens, a multimodal Jacobian lens mapping VLM intermediate states into a shared semantic space defined by the decoder vocabulary, testing global workspace theory (GWT) in multimodal AI. Key findings: concept info is readable in intermediate states and selectively accessed per task; counterfactuals show accessed concepts participate in downstream reasoning; two J-space metrics track cross-modal alignment better than hidden-state cosine, with reliable selective access emerging later than accessibility; and shared patterns hold across VLM families and sizes.

Original post →

More from Research

Research channel →