J-Space Surfaces a Model's Unspoken "Distracted" Thoughts

mlpowered · x · 2026-07-07

Researchers instructed the model to "simultaneously think about something irrelevant" while processing text and observed the J-space. The model complied without mentioning those irrelevant concepts in the generated tokens; yet, the J-space clearly surfaced them. This suggests J-space carries implicit thinking that the model leaves unexpressed in its output.

Related event: Anthropic Discovers Global Workspace Inside Claude(102 posts)→

Original post →

More from Research

Research channel →