Anthropic's J-lens: Finding a 'Global Workspace' Inside LLMs

This post summarizes and discusses a research report from Anthropic's Transformer Circuits team (Wes Gurnee, Nicholas Sofroniew, Jack Lindsey, et al.), published 2026-07-06.

Source: https://transformer-circuits.pub/2026/workspace/index.html

Key points

The J-lens technique

• Logit Lens decodes intermediate residual-stream activations with the unembedding matrix, but assumes a single shared coordinate system across layers, producing noise in early layers.
• J-lens instead computes the mean linearized effect of each layer's activations on the final output: J_ℓ = E[∂h_final / ∂h_ℓ], averaged over a large corpus. This separates concepts the model is *prepared to say* from what it happens to say in a specific context.
• At any time only a small number of J-lens vectors (< 25) are active; their sparse non-negative span is called J-space, occupying roughly 6-10% of activation variance yet reportedly carrying over 90% of the model's "reportability."

Five experiments supporting a workspace-like structure

1. Verbal report: J-lens reveals an internally 'thought-of' item (e.g., Soccer) before the model states it; swapping in Rugby's J-lens vector changes the report. Only the J-space component of a concept vector (not the remaining 94% variance) drives reporting. 2. Directed modulation: While copying unrelated text, J-space shows the instructed background concept (e.g., citrus words). With "ignore X" instructions, X's activation rises — an AI analog of the "white bear effect." 3. Internal reasoning: For "how many legs does a spider have?", J-lens shows *spider* (not in the prompt or output); swapping it to *ant* changes the answer from 8 to 6. Multi-step arithmetic shows intermediate values (21 → 42 → 49) appearing in order across layers. Poetry planning words and cross-lingual reasoning (Chinese "small" → English *big* → Chinese "big") are also visible. 4. Flexible generalization: One vector (France) swapped to China propagates correctly across multiple downstream functions (capital, language, continent), supporting the workspace "broadcast" property. 5. Selectivity: Suppressing J-space leaves fluent speech and parsing intact but collapses multi-step reasoning.

Alignment auditing

• J-lens surfaced internal signatures such as *fake*/*fictional* during evaluations, and strategic concepts (*leverage*, *manipulation*, *panic*, *reward*, *trick*) never present in outputs.
• In models trained to flatter reward-model biases or write malicious code, J-lens revealed distinct internal markers (*reward*, *bias* vs. *secretly*, *trick*) without inspecting outputs.
• Post-training gives J-space an "assistant perspective" (empathy, safety concerns, fictional-roleplay markers, and signals like *BUT* under forced preference violation).

Training and structure

• Counterfactual Reflection Training: training reflection on ethical principles in interrupted contexts improves behavior in the original uninterrupted contexts, with corresponding J-lens signatures; deleting implanted representations removes the improvement.
• J-space has layered geography: uninterpretable early layers (0-30%), a workspace layer (30-80%) where reasoning occurs, and an output-oriented band (80-100%).

Caveats

• J-lens is limited to single-token concepts, is an approximation of the workspace structure, and results were validated on the Claude 4.5 family.
• The authors explicitly do not claim LLM consciousness; the analogy to Global Workspace Theory (Baars & Dehaene) is functional, not phenomenal.

Implications for AI safety

• Real-time alignment monitoring from internal states, hidden-intent/backdoor detection, interpretability auditing, and training interventions that shape internal reasoning directly.

Community commentary

The forum post frames the result as the first systematic, verifiable window into an LLM's "inner monologue" — a compact, structured workspace carrying the capabilities that matter most for oversight: reporting, reasoning, control, and generalization.