Probing the "Conscious" Mind: Selective Access and Global Workspace in Vision-Language Models
A study of vision-language models finds that concepts can be present in internal representations before the model can reliably select them for a task. Its analysis tool compares concept accessibility across image and text pathways.
The paper introduces **MC-Lens**, which maps vision-encoder and language-decoder states into a shared semantic space and supports interventions on accessed concepts. Experiments find that concept information is readable internally, while reliable selective access develops later; accessed concepts also affect downstream reasoning.
For builders using vision-language models, output quality may depend on whether the model can select relevant visual concepts for a particular task, not only whether those concepts are represented. MC-Lens offers a research method for probing that distinction.
The paper introduces **MC-Lens**, which maps vision-encoder and language-decoder states into a shared semantic space and supports interventions on accessed concepts. Experiments find that concept information is readable internally, while reliable selective access develops later; accessed concepts also affect downstream reasoning. For builders using vision-language models, output quality may depend on whether the model can select relevant visual concepts for a particular task, not only whether those concepts are represented. MC-Lens offers a research method for probing that distinction. The findings describe controlled experiments across model families and sizes. The material does not establish that these measurements predict performance in deployed applications.
MC-Lens separates whether a vision-language model contains readable information about a concept from whether it can selectively access that concept for a task. The interventions also connect accessed concepts to downstream reasoning. This offers a way to probe a possible source of task-specific failure, but the controlled findings do not establish deployed performance.