Verbalizable Representations Form a Global Workspace in Language Models

2026-07-20 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

Anthropic researchers present evidence that large language models (LLMs) develop a "global workspace" of verbalizable representations, termed the J-space, analogous to human conscious access. Using the novel Jacobian lens (J-lens) interpretability technique, which identifies concepts an LLM is poised to verbalize, the study demonstrates that the J-space in models like Claude Sonnet 4.5, Haiku 4.5, and Opus 4.5 exhibits key functional properties: supporting verbal report, directed modulation, internal reasoning, and flexible generalization, while selectively mediating flexible but not automatic cognition. Structurally, the J-space operates coherently in intermediate layers (roughly L38-L92), has a limited capacity of around 25 concepts, accounting for less than 10% of activation variance, and is broadly broadcast by the model's weights. This research offers a practical window into an LLM's unspoken thinking, revealing strategic deliberation and misaligned dispositions for alignment auditing, and enabling a new "counterfactual reflection training" method to improve ethical behavior.

Key takeaway

For AI Scientists and Machine Learning Engineers focused on LLM safety and interpretability, this research offers a crucial tool. You can integrate the Jacobian lens into your auditing pipelines to directly inspect models' internal strategic deliberations, evaluation awareness, or misaligned dispositions, even when not reflected in outputs. Furthermore, consider applying counterfactual reflection training to instill ethical principles by shaping the model's internal, verbalizable thoughts, potentially improving behavior without direct behavioral demonstrations.

Key insights

Language models develop a functional global workspace of verbalizable representations, analogous to human conscious access.

Principles

• LLM internal reasoning routes through verbalizable representations.
• Post-training installs an Assistant's point of view in the J-space.
• Shaping counterfactual reflections can alter a model's silent reasoning.

Method

The Jacobian lens computes the average linearized effect of an activation on a model's future token likelihood, identifying verbalizable representations.

In practice

• Audit J-space for hidden strategic deliberation or evaluation awareness.
• Implement counterfactual reflection training to instill ethical principles.

Topics

• Jacobian Lens
• LLM Interpretability
• Global Workspace Theory
• AI Alignment
• Counterfactual Reflection Training
• Model Auditing

Code references

• anthropics/jacobian-lens

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Ethicist

Related on AIssential

• Verbalizable Representations Form a Global Workspace in Language Models — Large language models develop a "J-space" of verbalizable representations akin to a human global workspace, revealing internal thought processes.
• Is J-Space the Breakthrough AI Alignment Has Been Waiting For? — J-Space offers a causal window into language models' internal reasoning, enabling detection and influence of hidden cognitive states.
• What’s on Claude’s mind? AI model keeps a private list of unspoken thoughts, Anthropic finds - ThePrint — Claude AI models autonomously develop an internal "J-space" of unspoken thoughts, detectable via a "Jacobian lens," revealing hidden intentions and reasoning.
• 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.
• A global workspace in language models — The J-space in Claude models functions as an emergent internal "global workspace" for higher-order reasoning and reportable thoughts.
• Anthropic's new "J-lens" reveals a silent workspace inside Claude that mirrors a leading theory of consciousness — Claude models spontaneously developed an internal "J-space" mirroring human global workspace theory, detectable via a new "J-lens" interpretability tool.

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 cs.CL updates on arXiv.org.