arxivcs.CLcs.AIcs.LG2026-07-01
The State-Prediction Separation Hypothesis
Giovanni Monea, Nathan Godey, Kianté Brantley, Yoav Artzi
Transformers use the same forward computation stream to both predict the next token and store useful state for future token predictions. We formulate the \emph{state-prediction separation hypothesis}: disentangling the two roles yields better language modeling performance. We des…