This paper explores the notion that of all the information processed by the human brain, only a small fraction is consciously accessible—available for verbal report, deliberate control, and flexible reasoning. We propose that a similar functional distinction has emerged in large language models. Using a novel interpretability technique known as the Jacobian lens, we identify the representations a model can verbalize at any given moment, collectively termed the J-space.
The J-space exhibits functional properties characteristic of a global workspace: its contents can be reported, deliberately summoned and held; it is used to carry intermediate steps of silent reasoning, and can be passed as arguments to arbitrary downstream computations, while automatic processes like text parsing and routine inference proceed without them. Additionally, the J-space possesses structural signatures associated with conscious access, carrying coherent content only in an intermediate layer band, holding tens of concepts at a time, and being broadcast more widely by the model's weights than other representations. These traits render it a practical window into a model's unspoken thinking.
In alignment audits, the J-space reveals strategic deliberation, evaluation awareness, and misaligned dispositions that are not evident in the model's outputs. We find that post-training installs the Assistant's perspective in the workspace and introduces counterfactual reflection training, which enhances behavior by training what a model would say if interrupted and asked to reflect. These findings indicate that language models maintain a small, privileged set of representations with some functional hallmarks of conscious access, and decoding these representations illuminates ongoing cognitive processes.
Blogger's Review: This paper provides vital insights into the inner workings of language models, particularly through the in-depth analysis of J-space, advancing research on consciousness and representation in artificial intelligence.