Verbalizable Representations Form a Global Workspace in Language Models
A recent paper by contributors including Jack Lindsey from Anthropic presents evidence that large language models (LLMs) develop 'verbalizable representations' akin to a global workspace in humans. Using a technique called the Jacobian lens (J-lens), researchers identified a small, privileged set of internal representations that LLMs use for report, modulation, and flexible internal reasoning. These 'J-space' representations influence the model's internal thoughts and not just its output.
This research provides a novel interpretability technique, the J-lens, to understand LLM decision-making beyond just their outputs. It offers new avenues for auditing model safety, shaping internal thoughts through 'counterfactual reflection training,' and aligning AI behavior.
Learn one new AI thing every day.
Daily Deck sends you seven plain-English cards like this every morning. Free.
Start free