Stanford, Together AI debut NCM for 10–31× faster text generation

A new language-modeling framework from Stanford's Hazy Research group and Together AI reports text generation 10 to 31 times faster than strong diffusion baselines while using 4 to 5 times fewer training tokens.

The system, called Next-Concept Modeling (NCM), takes a different route from both token-by-token autoregression and iterative diffusion denoising. "NCM learns to represent text as a hierarchy of discrete codes, then generates one level at a time, predicting all codes within each level in parallel," the authors write in a Hazy Research blog post. The post opens on the research question directly: "Can we generate quality text in just a handful of parallel steps?"

The headline speedups split by task. XSum summarization matches the strongest diffusion baseline on ROUGE-L at 31× the throughput; LM1B unconditional generation runs 21× faster at better perplexity; OpenWebText runs 10× faster at comparable quality. On hard Sudoku puzzles with 30 clues, NCM solves 86.1% versus 81.3% for the DUO-FML+ diffusion baseline. Training uses 4.9× fewer tokens than diffusion on LM1B and 4× fewer on OpenWebText.

The comparisons are drawn against diffusion LMs "at similar parameter size," as lead author Hermann Kumbong put it in the project announcement, not against frontier autoregressive models at scale. The team lists Kumbong, Ishan S. Khare, Simran Arora, Azalia Mirhoseini and Christopher Ré among the authors; code is on GitHub and the paper is posted as arXiv:2610.14162.