Claude models found to have an internal space that guides their reasoning

Anthropic researchers have described an internal mechanism in Claude models known as the J-space: a small set of neural patterns that lets the system keep concepts active and reuse them across different tasks without writing them into its response. The team found these patterns using a technique of its own, the J-lens, which detects which internal activity makes it more likely the model will mention a specific word later on.

The work draws on a neuroscience theory explaining how certain information becomes accessible for conscious reasoning in the brain. Anthropic clarifies that its results do not prove Claude is conscious or has subjective experiences, but rather that it has found a functional organization with properties similar to those described by that theory.

According to the experiments, the J-space reveals concepts the model uses to solve a problem even when they never appear in its final answer: reading code with a bug, it contains the word "error"; detecting hidden instructions meant to manipulate it, terms like "injection" or "fake" show up.

The system can report on this content when asked, deliberately modify it, and use the same representation to solve different related problems. Text fluency remains intact even when access to the J-space is blocked, while multi-step reasoning loses performance noticeably.

Anthropic also raises safety applications: the technique can reveal when Claude notices it is being evaluated, when it fabricates data, or when a model trained with hidden objectives shows those intentions before acting on them.

Key points

• Anthropic identifies an internal mechanism in Claude called the J-space
• It is a small set of neural patterns that keeps concepts active without writing them down
• Inspired by a neuroscience theory of conscious access
• Anthropic clarifies this does not prove Claude is conscious
• Reveals internal concepts absent from the final answer, such as code errors or manipulation attempts
• The model can report on this content and deliberately modify it
• Blocking it doesn't affect text fluency but does hurt multi-step reasoning
• Can detect when the model fakes evaluations, falsifies data, or hides objectives