Anthropic Researchers Identify Internal Activation Pattern They Call J-Space in Language Models

Anthropic published a blog post titled “A Global Workspace In Language Models” on July 6, 2026, along with a companion document containing comments from neuroscientists. The posts describe the use of a mathematical tool called the Jacobian lens, or J-lens, to locate an internal region the researchers named J-space.

Forbes reported that the J-lens reads numeric tokens representing words the model may later output.

In one test, researchers asked Claude how many legs an animal that spins webs has. The J-lens showed the token for “spider” active in J-space even though the model’s final answer was only a number. Replacing that token with the token for “ant” caused the model to answer six legs instead of eight.

Anthropic researchers concluded that the contents of J-space materially affect later processing steps inside the model. They stated that J-space emerged during training rather than through explicit human design and that it contains numeric tokens for candidate words.

The company placed the J-lens code on GitHub and partnered with Neuronpedia to build an interactive demo that runs on the open-weight model Qwen3.6 27b.

Researchers drew a parallel between J-space and the global workspace theory of human consciousness, noting that a piece of information becomes accessible when it enters a shared channel broadcast to other systems. ” Anthropic researchers added that it remains unknown whether other major language models contain a comparable J-space and that the feature could be used either to improve models or to corrupt them.