Language models hide most of their actual computation in a geometric blind spot

Researchers studying how language models process information layer by layer found that the models split their work into two distinct parts: one visible and straightforward, one hidden and geometrically different. The hidden part is where the actual computation that changes the model's output lives — which means the standard ways of understanding how these models work are incomplete.

For years, AI researchers have assumed layer updates in language models follow predictable, uniform patterns. This paper shows they don't. The finding matters because it means current methods for probing, auditing, or controlling what happens inside language models are looking in the wrong place — they're measuring the visible layer and missing the separate computation that actually drives the model's behavior. This is the kind of structural gap that could explain why language models sometimes behave in ways their designers didn't anticipate.