Next-Concept Modeling (NCM) is a text-generation framework from Stanford's Hazy Research lab, described in an October 2026 paper and blog post. Instead of predicting one token at a time from left to right, NCM represents text as a hierarchy from coarse concepts to fine details and generates one level at a time, predicting every code within a level in parallel.
How Next-Concept Modeling works
- Stage 1, a multiscale tokenizer: a residual VQ tokenizer learns to turn text into a coarse-to-fine hierarchy of discrete codes. Early levels are short and capture broad structure; later levels are longer and add detail.
- Stage 2, a next-concept transformer: with the tokenizer frozen, a transformer learns to predict each level from the reconstruction of the earlier ones.
- Generation: the model fills one level at a time, all positions in parallel, and the tokenizer's decoder turns the finished hierarchy into text. The number of steps equals the number of levels, not the number of words.
Think of it as sketching first: the model starts with a rough idea of what it wants to say, and each level makes the text clearer.
Results the team reports
The researchers compare NCM mainly with diffusion language models, the leading alternative to token-by-token generation, which still need many sampling steps over the full sequence. Speed was measured as wall-clock time on a single NVIDIA H100 GPU.
- 21x faster than the strongest diffusion baseline on LM1B and 10x faster on OpenWebText, with better or comparable generative perplexity.
- Matches the strongest diffusion baseline on XSum summarization while running 31x faster.
- Solves 86.1% of hard Sudoku boards, against 81.3% for the strongest baseline in the comparison (FMLM+) and 1.0% for an autoregressive model.
- Uses roughly 4.9x and 4x fewer training tokens than diffusion baselines on LM1B and OpenWebText, counting tokenizer training.
What it does not show yet
These are research-scale models, and the comparisons are against diffusion language models rather than today's large production chatbots. The team lists open questions: how NCM scales with more parameters and data, whether training the tokenizer and generator jointly helps, and whether the model can learn to spend more computation on harder inputs. Weights and code are public on GitHub and Hugging Face.
For the basics of how today's left-to-right language models generate text, see our course on how ChatGPT and LLMs really work.
