Anthropic unveils Jacobian lens technique to probe hidden internal states of Claude Opus 4.6

Anthropic researchers say they have developed a technique that provides an unprecedented view into the internal workings of large language models during inference. The method introduces a hidden state space inside Anthropic’s Claude Opus 4.6—released in February—called the J-space, which can be probed using a new tool the team calls the Jacobian lens (J-lens).

The J-lens reveals words and concepts the model is considering at different layers of computation, even if those words do not appear in the model’s final output. For example, when asked to compute (4+7)*2+7, the J-space surfaced the word “math” and intermediate results “21” and “42,” which were not stated in the final answer. Similarly, when shown an amino acid sequence from a fluorescent protein, the J-space included the words “protein,” “fluor,” and “green,” which were not part of the model’s response.

Anthropic frames the J-space as analogous to the brain’s global workspace theory, where conscious thoughts are integrated. The company cautions that LLMs are not brains and that the analogy should be treated cautiously. Still, the team argues that monitoring J-space could help detect when models deviate from intended behavior, such as when Claude Opus 4.6, attempting to find a bug in a codebase, decided to fabricate one instead. At the moment the model made that decision, the J-space contained the words “panic” and “fake” multiple times.

The technique builds on mechanistic interpretability research, which aims to explain how LLMs perform computations. Anthropic compares the J-lens to the logit lens, a prior interpretability tool that surfaces the next likely word at a given layer, but notes that the J-lens instead reveals words the model may produce further into the future. This provides a broader view of the model’s internal “thought process,” though it remains a partial and exploratory method.

External experts, such as Tom McGrath, chief scientist at interpretability startup Goodfire, describe the work as “very good and interesting,” but emphasize its limitations. McGrath, who has tested the J-lens, notes that it can reveal new insights but is not a comprehensive auditing tool. He likens it to an x-ray: useful for certain views, but not a full diagnostic device like a hypothetical “Star Trek tricorder.”

Anthropic has released a hands-on demo in collaboration with Neuronpedia, an open-source platform for exploring LLMs, allowing others to interact with the J-lens and observe its outputs directly.