Within Claude’s language model, a small group of expressions, J-space, was found to hold concepts and multi-stage inferences that were not printed out, and experiments in reading and intervention showed that this group of expressions was causally involved in higher processing.
Internal spaces of words born during learning.
Anthropic identifies a small group of expressions with a special role within Claude’s language model.J-It’s called space.J-Space is said to have been formed spontaneously during model learning, not as part of the design.
Each pattern corresponds to a specific word, however, the activation of the pattern does not mean the output of the word and indicates that the concept appears in internal processing.
In Jacobian, we write internal activity into vocabulary.
To read out internal expressions, the Jacobian lens (J-lens) is used.This method requires an internal activity pattern for each word in the vocabulary that increases the likelihood that the word will come out in the future instead of now.
In the public reference implementation, the average input output Jacobian is used to linearise the residual flow of each layer to the bottom of the final layer. The output can be read as a word without having to separate the external classifier, so that the output can be converted into the order of the vocabulary token through the model’s own unembedding.
Tracking the unspoken calculations.
When the Jacobian lens is applied to multiple layers of Claude, the word appearing in the J-space can be tracked as it changes as processing progresses.Concepts that do not exist in the input or output are also covered, and words that respond to unspecified defects in the code and the processing of math problems appear inside.
In Anthropic’s experiment, the content of J-space was easier to linguize and report by Claude himself than by other internal expressions. The corresponding patterns were also activated when he was instructed to think of a particular object, and it was easier to adjust to the instructions than other expressions than J-space.
Intervention in the middle stage to verify the cause and effect relationship.
After solving the multi-stage problem, the middle stages that were not exported to the outside appeared in the correct order in J-space. This indicates that, apart from the chain of thought displayed, the process of reasoning can be observed at the stage of internal activity.
In addition to observed correlations, intervention experiments that manipulated internal patterns also confirmed their role.Anthropic says that the patterns in J-space mediated the results of multi-stage tasks causally and interfered with their use, although the normal interaction continued, the higher cognitive function was lost.
Sharing one concept into multiple processes.
The concept activated in J-space was not fixed on a single type of output and was flexibly used for different tasks.For example, after the expression for “France” was activated, separate information was extracted, such as the capital, currency, and the continent to which it belonged.
Anthropic has also obtained evidence that J-space is strongly connected to other parts of the neural network and has positioned it as a shared pathway similar to the global workspace of neuroscience by passing information through this connection to multiple processes.
Use the open-weights model.
The reference implementation of the Jacobian lens has been made public, and the open-weights decoder transformer allows you to learn, apply, and visualize the lens by layer and location.
The implementation provides a pathway to read pre-learned lenses, and a pathway to estimate the average Jacobian from your own prompt collection to make lenses. This allows you to apply the core reading method in Claude to another language model where the weight was published.