SciGroveBeta
Neuroscience

Verbalizable Representations Form a Global Workspace in Language Models

Wes Gurnee, Nicholas Sofroniew, Jack Lindsey

Featured July 29, 2026

Original Paper

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

Researchers found that large language models have a special internal "thinking space," called the J-space, where they hold ideas they could "say" if asked, much like our own conscious thoughts, allowing them to reason and plan.

In depth
The paper introduces the Jacobian lens (J-lens), a novel interpretability technique that identifies internal representations in large language models (LLMs) that are poised for verbal report. These "verbalizable" representations, collectively termed the J-space, exhibit functional and structural properties analogous to a human global workspace, enabling flexible reasoning and deliberate control over information.

Key Takeaways

  • 1
    The Jacobian lens is a new interpretability method that identifies internal LLM representations available for verbal report by measuring their causal effect on future token likelihoods.
  • 2
    The identified J-space functions as a global workspace, supporting internal reasoning, directed modulation, and flexible generalization, akin to conscious access in humans.
  • 3
    Counterfactual reflection training leverages the J-space to implant ethical principles, demonstrating a causal link between verbalizable concepts and silent reasoning, thereby improving model behavior.

Conceptual Flow

HIGH LEVEL
1
Methodology: How LLMs "Think" Out Loud

This shows how a special tool looks inside a computer brain to find the ideas it's ready to talk about, even if it doesn't say them yet.

Computer Brain's Inner Workings
Find Ready-to-Speak Ideas
List of Hidden Thoughts
2
Results: A Brain-like "Thinking Space" Emerges

The computer's hidden "thinking space" acts like a human's conscious mind, letting it plan, learn, and even reflect on its own actions.

Hidden Thoughts (J-space)
Enables Flexible Reasoning
Better Decisions
Self-Correction
Ethical Behavior

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.