A global workspace in language models
Reading
Paper audio
About this week
How can we identify and understand the internal “workspace” where large language models integrate information, reason, and broadcast concepts across their networks? Can mechanistic interpretability reveal whether LLMs develop functional equivalents of global information sharing found in cognitive architectures?
Anthropic’s Global Workspace research investigates whether large language models contain internal structures analogous to the global workspace theory of cognition—a framework proposing that intelligence emerges when specialized processes compete for access to a shared information space. By tracing how concepts and features propagate through model activations, the researchers identify patterns of information routing and representation sharing that resemble a distributed workspace.
Rather than treating models as opaque functions, this work attempts to map the internal communication pathways that support reasoning and abstraction. The findings suggest that advanced capabilities may emerge not from a single reasoning module, but from the interaction of many specialized representations that become globally available across the network. This provides a new lens for understanding scaling: as models grow, intelligence may depend not only on increasing parameter count, but on developing richer internal mechanisms for information integration, coordination, and reuse.
Join us at CASI for discussion at 8 pm, and (optional) quiet reading from 7 pm.
