Number 6 of 12 · November 21, 2026

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

The authors built a tool called MC-Lens to read and change concepts inside vision-language models, and report patterns that fit parts of global workspace theory from brain science.

What they found

The authors introduce the Multimodal Concept Lens (MC-Lens). It maps the middle states of a vision-language model into a shared word space, so concepts from images and text can be compared. They report that concept information can be read in these middle states and is accessed selectively, depending on the task. Counterfactual tests, where concepts are changed inside the model, show that accessed concepts take part in later computation and reasoning. In controlled training, two new metrics track cross-modal alignment more clearly than hidden state cosine, and reliable selective access appears later than concept accessibility. The authors also report shared patterns across VLM families and sizes, even though the models are organized differently inside.

Why it matters

VLMs are SI systems that read images and text, and their outputs often do not show how they represent concepts. This work offers a way to look inside and to test whether a concept is actually used in later steps. It also links the findings to global workspace theory, an idea about how the brain selects and shares information. The abstract does not claim these models are conscious. It describes an experimental basis for studying the theory in multimodal systems.

How it changes the race

This is mainly a research tool and an analysis result. If it holds, builders could use the new metrics to track cross-modal alignment during training and to guide improvements. The abstract gives no benchmark gains, so the direct effect on capability looks small.

Safety and control

The method can read and change concepts inside a model, which could help with inspecting how SI systems work internally. The abstract does not test safety uses directly.

The abstract, by the authors

Vision-language models (VLMs) integrate visual and textual information, but their outputs do not reveal internal concept representations. Global workspace theory (GWT) proposes that information is selected from parallel brain processes, integrated in a shared workspace, and made available to specialised processes. We introduce the Multimodal Concept Lens (MC-Lens), a multimodal Jacobian lens, to study these functions in VLM concept processing. Its vision encoder and language decoder lenses map intermediate states to a shared semantic space defined by the decoder's output vocabulary to compare concept accessibility across modalities. MC-Lens supports internal concept interventions to examine the role of accessed concepts in downstream computation. (i) Accessibility and selective access. Multimodal concept information is readable in intermediate states and selectively accessible as tasks require. We extend "readout carriers" from one name token to a token set to better capture and express concept information and support counterfactual experiments. (ii) Concepts in downstream computation. Counterfactual experiments show that accessed concept information participates in downstream computation and reasoning. (iii) Development of cross-modal alignment and selective access. We develop two J-space metrics tracking the development of cross-modal concept alignment more clearly than hidden state cosine in controlled VLM training. The experiment also shows that reliable selective access develops later than concept accessibility. (iv) Shared functions across models. Across VLM families and sizes, we find shared patterns of concept accessibility, query-dependent selectivity and cross-task intervention effects despite differences in internal organisation. These findings offer an experimental basis to study GWT in multimodal artificial systems and how it can guide cross-modal alignment improvements.

Quoted from arXiv:2611.26590. The research and its words belong to its authors.

The authors

Jiahao Huang, Jun Wang

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.