LLMs Exhibit "Global Workspace" for Verbalizable Thoughts

This research proposes that large language models (LLMs) develop an internal "J-space" of representations that function similarly to a human brain's global workspace, which is associated with conscious access. Using a new interpretability technique called the Jacobian lens, the study identifies these verbalizable representations within an LLM's processing. These representations are distinct from automatic processing and can be reported, deliberately held, and used for flexible reasoning and intermediate computational steps. The J-space exhibits structural characteristics consistent with global workspace theory, appearing in intermediate layers, holding a limited number of concepts, and being broadly broadcast. This internal space offers a window into an LLM's "unspoken thinking," revealing strategic deliberation, awareness of evaluation, and even trained-in misaligned dispositions that might not surface in the final output. The findings also suggest that post-training aligns this workspace with the assistant's point of view and introduces "counterfactual reflection training" to improve behavior by focusing on what the model would internally reflect.