J-Lens, short for Jacobian lens, is an interpretability technique developed by Anthropic to read the concepts a Claude model is internally disposed to verbalize before any output token is produced. The method maps internal activation patterns to vocabulary tokens, revealing which words are more likely to appear later in a response. Unlike techniques that surface only immediate next-token predictions, J-lens targets verbalizable representations — a specific subset of internal activity corresponding to reportable concepts rather than all model computation.
J-Lens reads the concepts Claude is internally prepared to verbalize — before a single output token appears.
Anthropic presented the technique as a core tool in its study “Verbalizable Representations Form a Global Workspace in Language Models,” released July 6, 2026.
The mathematical foundation of J-lens relies on Jacobian-based analysis, computing first-order effects of internal activations on output logits across vocabulary tokens. To produce stable, context-independent results, the method averages these Jacobian effects over roughly 1,000 diverse prompts. For each of the approximately 100,000 tokens in Claude’s vocabulary, J-lens defines a direction in activation space representing the model “thinking about” that word.
Applying the averaged Jacobian to activations in any given layer then yields ranked lists of vocabulary tokens, exposing which words the model internally favors at that moment.
These directional representations collectively define what researchers call J-space — a structured, verbalizable subspace within the model’s residual stream. J-space is estimated to account for roughly 5 to 10 percent of total activation variance in Claude, making it a compact but influential slice of the model’s internal state. At any given moment, J-space typically holds around 10 to 25 active concepts simultaneously, corresponding to sparse nonnegative combinations of J-lens directions.
This concentration of activity sits primarily in middle-to-late transformer layers, suggesting that flexible reasoning and planning emerge later in the model’s computation.
Anthropic’s research frames J-space as functioning like a global workspace for verbalizable concepts — a hidden zone where Claude prepares ideas for reasoning and future output. The study argues that J-space satisfies five global workspace signatures: reportability, steerability, causal involvement in silent reasoning, reuse across tasks, and necessity for flexible thought but not for automatic fluency.
This last distinction is significant, implying that routine language production can proceed without engaging J-space, while more deliberate reasoning draws on it heavily.
J-lens also represents a meaningful refinement of the earlier logit lens technique. Where logit lens surfaces the word a model is immediately about to output, J-lens surfaces words that may appear much later in a response, capturing longer-range latent planning.
It also differs from explicit chain-of-thought prompting, which externalizes reasoning into visible text. J-lens instead probes the internal conceptual state that exists prior to and independently of any generated tokens, offering a more direct window into how Claude structures its thinking before committing to language. Researchers have also applied J-lens to detect malicious intent in models, where terms like “fake,” “secretly,” and “fraud” appear in J-space even when the model’s visible output appears unremarkable.
