A global workspace in language models

This research summary presents evidence that the Claude language model develops a "J-space": a small set of internal activation patterns linked to particular words that act as a silent workspace for reportable thoughts and multi-step reasoning. The authors describe experiments (reading, intervening, and swapping these patterns) showing the J-space's role in internal reasoning, monitoring for misbehavior (fabrication, manipulation, evaluation-awareness), implications for model safety and training (counterfactual reflection), and philosophical connections to theories of conscious access.