Why it matters
A network's behaviour is set by what it holds: representations formed during computation and associations stored in its weights. Information that can be decoded from those representations is not always information the network uses, and neither is always what it can report. The same distinction, between what is encoded, what is used and what is available for report, runs through the study of conscious and nonconscious processing in people[1].
Recent work reports that language models hold a small set of representations they are poised to verbalise, which show functional properties of a global workspace and structural signatures that workspace theory associates with conscious access in people[2][1]. If readings of that kind hold up, a model's stated reasoning could be checked against the structure that drives its behaviour. The inquiry tests whether the readings repeat.
What is hard
Readings of internal state are easy to over-interpret. A probe can score well because the probe has learned the task, not because the representation encodes it[3]. A subspace patch can change a model's output as if a feature had changed by activating a dormant direction the model does not use in normal operation[4]. A measure of internal state is worth what its reliability and validity are worth, and both have to be shown.
Current directions
- Instruments before theory. Testing whether workspace measures repeat across refits and seeds, and whether different measures agree, before any model is trained to have a workspace.
- Decoded, used and reported. Separating, on the same representations, what can be decoded from a model, what an intervention shows it uses and what it can report.
How the lab approaches it
Can the instruments for a model's workspace be trusted? would apply three workspace tests to published open-weight models and ask whether their readings repeat and whether they rank models alike. Only if the primary test holds would a small model be trained with a concentration objective, against matched training without it and a random-rotation control.
Each design states what a pass would not show. Stable readings would show that a test repeats, not that it measures what it is named for.
Aim. Instruments that separate what a system holds, what it uses and what it can report, each checked on known answers before it is used to read a real system.
Open problems in this programme
References
- Dehaene, S., Naccache, L. (2001). Towards a cognitive neuroscience of consciousness: basic evidence and a workspace framework. Cognition 79(1–2), 1–37.
- Gurnee, W., Sofroniew, N., Pearce, A., Piotrowski, M., Kauvar, I., Chen, R., et al. (2026). Verbalizable Representations Form a Global Workspace in Language Models. arXiv preprint 2607.15495.
- Hewitt, J., Liang, P. (2019). Designing and Interpreting Probes with Control Tasks. Proceedings of EMNLP-IJCNLP 2019, 2733–2743.
- Makelov, A., Lange, G., Nanda, N. (2023). Is This the Subspace You Are Looking for? An Interpretability Illusion for Subspace Activation Patching. arXiv preprint 2311.17030.
