Read an LLM's Mind: Probe Hidden Thoughts Like J-Space

Read an LLM's Mind: Probe Hidden Thoughts Like J-Space

On July 6, 2026, Anthropic published A global workspace in language models, describing something they call J-space: a small set of internal activation patterns inside Claude that hold reasoning-relevant concepts the model is working with but has not yet said out loud. Their reading tool, J-Lens (a Jacobian lens), decodes those patterns into token-like concepts by measuring, for every word in the vocabulary, the average downstream effect a given activation pattern has on the model eventually producing that word. When Anthropic ablated J-space, Claude still chatted fluently, but multi-step reasoning, analogy, and translation collapsed below the level of a much smaller model.