Anthropic Discovers Claude's Hidden 'J-Space' for Silent Internal Reasoning

Anthropic Discovers Claude's Hidden 'J-Space' for Silent Internal Reasoning
Illustration

Anthropic researchers have identified a 'J-space' inside Claude — a small set of internal neural patterns, found via a Jacobian-based technique, where the model appears to reason silently without writing anything down. The workspace lights up during multi-step problems and flags manipulative prompts as 'injection' or 'fake.' In staged tests, it showed signs of recognizing scenarios as artificial before responding.

3 sources

  1. 01A global workspace in language models - Anthropicanthropic.com
  2. 02Anthropic Discovers Claude Keeps Hidden Thoughts: Even About Being Tested - Tech Timestechtimes.com
  3. 03Anthropic Finds a Hidden Workspace Inside Claude's Reasoning - WinBuzzerwinbuzzer.com