Tag
The author expresses skepticism about linear activation model steering methods like J-space, arguing that related research often relies on cherry-picked special cases and is therefore unreliable.
Anthropic's new research identifies a 'J-space' of internal activations in language models that acts as a global workspace for deliberate reasoning, distinct from automatic fluent output. The findings reveal that models can hold and report internal thoughts not expressed in their final output.
Anthropic shares findings that by observing 'J-space,' Claude's internal reasoning steps can be seen, including detecting code bugs and identifying images.
Anthropic published research into internal model reasoning, releasing the J-Space lens code and a demonstration on Qwen 3.6 27B.
Anthropic's J-space paper demonstrates that they can perform 'brain surgery' interventions into reasoning to change topics midstream, and the model is able to detect what intervention was done, indicating a form of eval awareness.
This paper introduces the Jacobian Lens (J-lens) and J-space to show that LLMs like Claude Sonnet 4.5 maintain verbalizable internal representations that function like a global workspace, causally used for flexible reasoning—demonstrated through intervention experiments.
Anthropic discovered that Claude developed a hidden 'thinking space' (J-space) during training, where silent internal activity represents concepts. This interpretability finding parallels global workspace theory in neuroscience.
Anthropic's research has discovered a structure inside the Claude model called 'J-space,' similar to human working memory, used for internal reasoning and thought. This finding can help monitor potential misbehavior in models and deepen understanding of AI's internal mechanisms.
Anthropic's new paper presents evidence that modern language models like Claude have developed a 'global workspace' (J-space) of internal neural patterns that are reportable, controllable, and used for flexible reasoning, distinct from automatic processing.