The State-Prediction Separation Hypothesis

[Submitted on 1 Jul 2026]

View PDF HTML (experimental)

Abstract: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 design a Transformer variant that uses two computation streams to separate the two functions, and conduct pretraining experiments across various scales. Our experiments show that state-prediction separation consistently offers better data and compute efficiencies, improving validation loss and outperforming standard Transformers by 2–3 percentage points on average on downstream tasks. We also conduct extensive empirical analysis that rules out potential confounders and demonstrates the fundamental difference in the gradients our design entails.

Submission history

From: Giovanni Monea [view email]
[v1] Wed, 1 Jul 2026 17:55:09 UTC (1,510 KB)