Autoregressive vs Diffusion Training: The Data-Efficiency Math Behind Non-Causal Architectures

mgostIH · x · 2026-09-11

ThePremiseOfIt explains autoregression's core training advantage: a length-T context yields T input-output pairs, so training scales with total tokens rather than total contexts; non-causal training yields only one pair per context, far less efficient — which is why diffusion models are hard to train and block diffusion took off (T//n pairs, but capped generation length n).

mgostIH counters that papers suggest diffusion models are more data-efficient, needing roughly half the tokens on average under uniform prefix/suffix splitting. The thread also touches on whether DeepSeek's V4.1 architecture could have used a non-causal encoder+decoder, and Premise says solving this is part of their work.

Related event: Debate Flares Over Data Efficiency of Diffusion vs Autoregressive Models(2 posts)→

Original post →

More from Research

Research channel →