Verbalizable Representations Form a Global Workspace in Language Models

2026-07-16 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A new interpretability technique, the Jacobian lens, identifies "verbalizable representations" within large language models, termed the J-space. This J-space functions as a "global workspace," analogous to human conscious access, enabling models to report, summon, and hold content, perform silent reasoning, and pass arguments to downstream computations. Structurally, J-space carries coherent content in an intermediate band of layers, holds tens of concepts, and is widely broadcast. This window into a model's "unspoken thinking" reveals strategic deliberation, evaluation awareness, and misaligned dispositions not evident in direct outputs. The research notes that post-training installs the Assistant's point of view in this workspace and introduces counterfactual reflection training, which improves model behavior by focusing on what the model would reflect if interrupted.

Key takeaway

For AI Scientists and NLP Engineers focused on model alignment and interpretability, this research offers a new avenue for understanding internal model states. You should consider integrating Jacobian lens techniques into your audit processes to uncover "unspoken thinking," including potential misaligned dispositions, before they manifest in outputs. Furthermore, explore counterfactual reflection training as a method to directly improve model behavior by shaping its internal reflective capacity.

Key insights

Large language models develop a "J-space" of verbalizable representations akin to a human global workspace, revealing internal thought processes.

Principles

• LLMs form a functional global workspace.
• Unspoken thinking can reveal misaligned dispositions.
• Post-training shapes the model's internal perspective.

Method

The Jacobian lens technique identifies verbalizable representations (J-space) by analyzing a model's readiness to report content. Counterfactual reflection training improves behavior by training on hypothetical internal reflections.

In practice

• Use J-space for alignment audits.
• Train models using counterfactual reflection.

Topics

• Language Models
• Model Interpretability
• Global Workspace Theory
• AI Alignment
• Counterfactual Reflection Training
• Jacobian Lens

Best for: Research Scientist, AI Scientist, NLP Engineer, AI Ethicist

Related on AIssential

• No Space Like J-Space — The J-space is a functional global workspace in LLMs, containing verbalizable representations crucial for internal reasoning and output determination.
• Vernier: Probing Representational Misalignment Behind Lexical Gaps in Causal Reasoning — Lexical gaps in causal reasoning by LMs stem from representational misalignment, not information loss.
• Anthropic J-lens reveals hidden workspace inside Claude — Anthropic's Claude models spontaneously developed an internal "J-space" for reasoning, detectable via the Jacobian lens, impacting AI safety.
• Towards surfacing model algorithms with meta-tokens in the J-Space — J-lens can surface "meta-tokens" in LLMs, revealing internal algorithms and computational states.
• Verbalizable Representations Form a Global Workspace in Language Models — Language models develop a functional global workspace of verbalizable representations, analogous to human conscious access.
• An AI lab says chatbots have what may be a key feature of consciousness. Are they right? And what now? — Anthropic's Claude demonstrates an internal "J-space" akin to a global workspace, fueling debate on artificial consciousness.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.