MC-Lens Paper Reads VLM States, Shows Selective Concept Access
TL;DR
- The authors introduce MC-Lens, pairing a vision encoder lens (V-Lens) and language decoder lens (L-Lens) built on averaged Jacobian matrices.
- The smallest greedy concept set, J@5, redirects nearly every one-hop VLM answer, while the broader word-form set F_≤3 redirects fewer than half.
- The authors use 'conscious' strictly as a functional label and explicitly do not claim the models possess consciousness or subjective experience.
Swap as few as five concept directions in a vision-language model's decoder and it will redirect nearly every one-hop answer. That is the headline intervention result in a new preprint on Hugging Face introducing MC-Lens, a Jacobian-based interpretability method for VLMs.
The authors describe it as "a multimodal Jacobian lens that combines a vision encoder lens (V-Lens) and a language decoder lens (L-Lens)." V-Lens reads concept information from intermediate vision-encoder states before the image is combined with the text input; L-Lens reads concepts from decoder states during visual-textual integration. On their counterfactual subspace-replacement intervention, they report that "the smallest greedy set, J@5, redirects nearly every one-hop answer, whereas the broadest word-form set, F_≤3, redirects fewer than half."
The paper frames the findings through global workspace theory: that concept information is readable in intermediate states, drives selective access depending on the text query, and "is not merely shown on a 'display board', but also participates in downstream computation and reasoning." The authors also report a training-time asymmetry — accessibility arrives early in alignment training, reliable selectivity later.
The disclaimer is unusually explicit for an interpretability paper: "We do not claim that these models possess consciousness or subjective experience." It lands in a busy stretch of multimodal interpretability work; our multimodal tracker logged 124 stories in the last 90 days.
Originally reported by huggingface.co
Read the original article →Original headline: MC-Lens Paper Finds Vision-Language Models Show Selective Access and Global-Workspace Patterns