Vision language models reveal how concepts are selectively accessed and shared

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

Artificial IntelligenceComputer Vision and Pattern RecognitionMachine Learning

Summary

Vision-language models combine pictures and words but it’s hard to see how they understand concepts inside their ‘mind.’ The authors created a tool called MC-Lens to peek inside these models and see how different ideas are accessed and shared between vision and language parts. They found that these models selectively focus on relevant ideas depending on the task, and those ideas influence what the model outputs next. Their work also shows how these abilities develop during training and work similarly across different models.

What this means in practice

  • For ai software engineers: Design debugging methods to inspect and intervene on internal concept processing in vision-language systems during model development.
  • For multimodal application developers: Improve cross-modal alignment to build more reliable and interpretable systems that jointly understand images and text.

Authors

Jiahao Huang, Jun Wang

Abstract

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.