Apple tests flow-based language models as an alternative to token-by-token generation

Apple researchers have scaled a method called categorical flow maps, testing whether language models can generate text through a continuous process instead of predicting one token after another.

Most modern language models are autoregressive: they produce the next token, then repeat the step until the answer is complete. Apple’s research explores a different route borrowed from diffusion and flow-matching systems, which are already common in image and audio generation. The work treats discrete text as a continuous generation problem before mapping it back into categories.

The practical appeal is speed and control. The paper says categorical flow maps can support accelerated sampling and “tilting,” a technique for steering generation, while staying competitive with existing approaches. That does not make the method a replacement for today’s large language models yet; it is still research, and the hardest test is whether it can scale reliably on real language tasks.

For developers and researchers, the update matters because it widens the set of credible architectures for language modeling. If the approach continues to improve, future text systems may not have to inherit every limitation of token-by-token decoding.