← News

Models & toolsSources: arXiv · 20 Nov 2026

MC-Lens probes concept access inside vision-language models

Researchers introduce a diagnostic tool that reads internal concept states in vision-language models and tests whether those states drive model reasoning.

In plain words

Vision-language models combine images and text, but their internal workings are hard to inspect. A research team built MC-Lens, a tool that maps a model's intermediate processing states onto the words it can output, making internal concepts readable. They report that models selectively surface concepts as tasks demand, and that this selectivity develops later than basic concept readability.

What was announced

The authors frame MC-Lens against global workspace theory, which holds that cognition selects information from parallel processes and integrates it in a shared workspace. Their tool runs two lenses, one over the vision encoder and one over the language decoder, and projects intermediate states into a shared semantic space defined by the decoder's output vocabulary. This lets them compare how accessible a given concept is from each modality. They also widen the readout from a single name token to a token set, which they say captures concept information more fully and enables counterfactual experiments.

The reported findings are threefold. Counterfactual interventions on internal concept states suggest that accessed concepts participate in downstream computation and reasoning rather than sitting inert. Two new J-space metrics tracked cross-modal concept alignment during controlled training more clearly than hidden-state cosine similarity, and showed that reliable selective access emerges later in training than concept accessibility itself. Across VLM families and sizes, the authors report shared patterns of accessibility, query-dependent selectivity and intervention effects despite differing internal organisation, offering an experimental basis for studying global workspace ideas in multimodal AI systems.

What it does not show

  • Evidence is limited to the paper's abstract; full methods, datasets and results are not available in the supplied material.
  • Findings such as shared cross-model patterns and intervention effects are the authors' own claims and have not been independently confirmed.

Sources

Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 21 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.