MC-Lens probes selective access and global workspace functions inside vision-language models
Jiahao Huang · hf · 2026-11-20
Imperial College London and UCL researchers introduce MC-Lens, a multimodal Jacobian lens mapping VLM intermediate states into a shared semantic space defined by the decoder vocabulary, testing global workspace theory (GWT) in multimodal AI. Key findings: concept info is readable in intermediate states and selectively accessed per task; counterfactuals show accessed concepts participate in downstream reasoning; two J-space metrics track cross-modal alignment better than hidden-state cosine, with reliable selective access emerging later than accessibility; and shared patterns hold across VLM families and sizes.
More from Research
- Team claims Codex cracked 35 open Erdős problems in 2 months, some Lean-verified — IgorCarron · 2026-11-20
- ForgeSystem paper proposes AI agents forging demand-specific software — Div_pradeep · 2026-11-20
- Accord: open-source dense metric mapping from a single camera and IMU — BE2R · 2026-11-20
- AltRA-Test: ~100 samples suffice to decide if LLM judges can replace humans — Technion · 2026-11-20
- TwinD hits 95.6% AUROC on face synthesis deepfakes, beating prior detectors — Andrii Yermakov · 2026-11-20
- Watch a single artificial neuron compute: weights, bias, and one activation function — TinfoilTricorn · 2026-11-20