Do Transformers Implement Global Workspace Dynamics?
Saman Rahbar
Zenodo (CERN European Organization for Nuclear Research) March 7, 2026 preprint DOI: 10.5281/zenodo.18907011 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Theoretical or philosophical paper |
|---|---|
| Key points | Argues that transformer language models may instantiate a functional analog to Global Workspace Theory through four candidate structural mappings, but contends that two critical disanalogies—the absence of temporal recurrence and the absence of recurrent broadcast—undermine the analogy, and proposes a falsifiable experimental program to test it. |
Abstract
This paper examines whether transformer language models instantiate a functional analog to Global Workspace Theory (GWT). Through a structured literature review spanning mechanistic interpretability, sparse autoencoders, consciousness theory, and brain-transformer alignment studies (through 2025), four candidate structural mappings are evaluated: the residual stream as shared workspace, attention heads as broadcast specialists, superposition as unconscious processing substrate, and induction-head phase transitions as ignition dynamics. Two critical disanalogies are identified — absence of temporal recurrence and absence of recurrent broadcast — and a falsifiable experimental programme is specified.