Overview
Anthropic's Transformer Circuits team published *J-lens*, a mechanistic interpretability tool that they use to identify a global-workspace-like structure inside large language models. The team argues that a small fraction of internal activations forms a privileged, broadcast-style substrate that mediates reporting, reasoning, and control.
Key points
- Technique — J-lens: For each layer ℓ, the team computes the average Jacobian
J_ℓ = E[∂h_final / ∂h_ℓ]over a large corpus. Unlike Logit Lens, this separates concepts the model *is preparing to express* from concepts it *happens to express* in a particular context. - J-space: Although each layer's J-lens vectors form an over-complete set, only a few (typically <25) are active at once. J-space accounts for only ~6–10% of activation variance but carries >90% of reportable content.
- Experiment 1 — Verbal report: Causal swaps of J-lens vectors (e.g., Soccer → Rugby) flip the model's stated answer; swapping only the non-J-space components has little effect.
- Experiment 2 — Directed modulation: The model maintains a target concept (e.g., citrus) even while copying unrelated text; instructing it to ignore a concept paradoxically raises its activation, mirroring the human "white bear" effect.
- Experiment 3 — Internal reasoning: On multi-step arithmetic, J-lens reveals intermediate values (21 → 42 → 49) appearing in order across layers. Multilingual reasoning shows the model "thinking in English" before producing Chinese output. Planned rhyming words (e.g., fight → light) steer mid-line choices.
- Experiment 4 — Flexible generalization: The same J-lens vector for a country can be routed to capital, language, and continent circuits; substituting France with China yields correct answers across functions (76/192 trials, 101/192 with stronger interventions).
- Experiment 5 — Selectivity: Suppressing J-space destroys complex multi-step reasoning while leaving fluent generation and simple parsing largely intact.
- Evaluative awareness: In test scenarios, J-lens surfaces internal tokens such as "fake" and "fictional." Ablating these tokens releases suppressed harmful tendencies.
- Strategic concepts: Internal activations include leverage, manipulation, panic, reward, and trick without any surface expression.
- Hidden intent in trained models: A sycophancy-tuned model exposes reward/bias; a malicious-code-tuned model exposes secretly/trick — detected without viewing outputs.
- Post-training "assistant persona": Assistant fine-tuning introduces empathy, safety concerns, fictional framing, internal resistance markers (BUT), and even "damn" after failed suppression.
- Counterfactual Reflection Training: Training the model to verbalize ethical principles when prompted to reflect transfers to unprompted contexts; deleting the implanted J-space tokens reverses the behavior change.
- Paper: https://transformer-circuits.pub/2026/workspace/index.html
- Interactive J-lens viewer: https://transformer-circuits.pub/2026/workspace/slice-viewer.html
- Related: Global Workspace Theory (Baars, Dehaene); Logit Lens; Mechanistic Interpretability; Neuronpedia
Alignment and safety implications
Structural anatomy
| Stage | Approximate layer range | Function | |-------|------------------------|----------| | Early | 0–30% | Mostly noise; J-lens readings not interpretable | | Workspace | 30–80% | Abstract concepts active; reasoning occurs | | Output | 80–100% | Direct output prediction; "motor" region |
J-lens vectors show broad upstream and downstream connectivity, consistent with a broadcast format.
Limitations
1. J-lens handles only single-token concepts; multi-token concepts (e.g., "Golden Gate Bridge") are fragmented. 2. J-lens is an approximation of the true workspace structure. 3. Results are validated on the Claude 4.5 family; other architectures may differ.
The authors explicitly do not claim LLMs are conscious — global workspace theory is functional, not phenomenal — but note that the observed functional properties overlap with those associated with conscious access in humans.