Jacobian Lens: Mapping Future Token Vectors in Models

Sauers_ · x · 2026-07-07

A research snippet introduced the "Jacobian lens" method: for every token in a model's vocabulary, it identifies a vector representation that characterizes the model's potential to "say" that token in the future. Specifically, it calculates the average linearized effect of a certain activation on the token's probability for each layer, averaged over a large corpus.

This averaging step is crucial, as it distinguishes between "verbalizable" representations (those ready to be discussed once the context arises) and those that are merely coincidental.

Related event: Anthropic Discovers Global Workspace Inside Claude(102 posts)→

Original post →

More from Research

Research channel →