Stanford's NCM writes text coarse to fine
Why it matters
Instead of one token at a time, Stanford Hazy Research's Next-Concept Modeling generates a hierarchy of codes level by level, predicting every code in a level in parallel. Against the strongest diffusion LMs at matched size on one H100: 21x faster on LM1B with better quality, 10x on OpenWebText at comparable quality, and the same ROUGE-L on XSum at 31x faster; 86.1% vs 81.3% on hard Sudoku; about 4-5x fewer training tokens. Early research: ~100M-parameter models, scaling tested to 725M.
Read the source: Stanford Hazy Research (opens in a new tab)
All hits from 19 Oct 2026More ResearchGet the daily hits by email