Scaling Categorical Flow Maps
Apple's research team shows flow matching can match autoregressive LMs at scale—opening a new path to faster language generation.
Why it matters
Continuous diffusion and flow matching represent a fundamental alternative to autoregressive decoding for language models. If CFMs scale competitively, they unlock accelerated sampling and other advantages currently unavailable to transformer-based LMs—reshaping how practitioners think about model architecture tradeoffs.
The key facts
6 to knowCategorical Flow Maps (CFMs) enable discrete data generation via continuous flow matching
Flow matching shows competitive performance vs. autoregressive language models
Accelerated sampling demonstrated; avoids sequential decoding bottleneck
Apple ML research (first-party, peer-reviewed venue signal)
Diffusion/flow matching bridges continuous and discrete modalities for LM
Published Aug 2026; recent research milestone
Go to the source
Apple Machine Learningmachinelearning.apple.com
Publisher excerpt: Continuous diffusion and flow matching models could represent a powerful alternative to autoregressive approaches for language modelling (LM), as they unlock a host of advantages currently reserved for continuous modalities, including accelerated sampling and tilting. Recently, several works have…