Stanford's Next-Concept Modeling generates coarse to fine, in parallel

On 14 October 2026, Hermann Kumbong posted Next-Concept Modeling, and the Stanford AI Lab account followed the same evening. The paper they point to was already on arXiv. Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation is dated 9 October 2026 (arXiv:2610.14162v1). The authors are Hermann Kumbong, Ishan S. Khare, Iddah Mlauzi, Andrew Woen, Charles Ding, E. Kelly Buchanan, Simran Arora, Azalia Mirhoseini, and Christopher Ré, all listed with Stanford computer-science affiliations. The paper marks Kumbong and Khare as equal contributors. Kumbong’s post also tags Hazy Research. The code lives at HazyResearch/ncm.

What the method is

The abstract describes two stages. A multiscale residual VQ-VAE encodes text into a hierarchy of discrete codes, coarse to fine. A transformer then generates one level at a time and predicts every code inside a level in parallel. The repository README puts the same idea in one line: one forward evaluation per level, because positions within a level are generated together.

That is a research architecture for parallel text generation. It is not a new hosted chat model, and the paper does not claim a comparison with frontier autoregressive systems such as the current large commercial models. The baselines it names are diffusion language models, flow-based and hybrid models, and, on Sudoku, autoregressive models of a similar experimental scale.

What the authors report

These figures are the authors’ own. They were not rerun here.

The abstract says that at matched parameter scales, NCM gets better generative perplexity than the strongest diffusion baseline on LM1B while running 21 times faster, is competitive with the strongest diffusion baseline on OpenWebText while running 10 times faster, and matches the strongest diffusion baseline’s ROUGE-L on XSum while running 31 times faster. The body names those speedups more tightly: on LM1B, 21 times faster than DUO at better generative perplexity; on XSum, 31 times faster than AURORA-LM at comparable ROUGE. A summary line in the introduction says 10 to 31 times faster generation at comparable or better quality. The README’s highlight line says 9 to 31 times. The paper’s range is the one to quote.

Conditional generation in the paper is XSum summarization and WMT14 German-to-English translation. On that pair the paper says NCM uses about 6 to 16 times fewer function evaluations than unguided baselines, with the best summarization scores and a translation result that trails only the strongest non-autoregressive baseline, E2D2, while using 5.8 times fewer function evaluations. The explicit 31 times wall-clock line is attached to XSum, not to a separate WMT timing. Kumbong’s post groups summarization and machine translation under “up to 31 times faster.” The paper’s 31 times sentence is the summarization one.

On Sudoku, the abstract says NCM outperforms the autoregressive and diffusion baselines across difficulty levels and uses fewer function evaluations. The introduction also says the model reaches competitive quality with about 4 to 5 times fewer training tokens than strong diffusion baselines on LM1B and OpenWebText, counting tokenizer and generator training.

What is public, and under which license

HazyResearch/ncm is public. GitHub records the repository as created 11 October 2026. The LICENSE file is the MIT license, copyright 2025 Hermann Kumbong. The README says the same.

Weights are a set of tokenizer and generator checkpoints, not one general chat model. The checkpoint card lists folders for XSum, WMT14, Sudoku (easy, medium, hard), LM1B (small, medium, large), TinyStories, and OpenWebText. The card’s frontmatter says license: mit. Hugging Face records the repo as created 13 October 2026 at 12:15 UTC and last modified 14 October at 05:29 UTC. The companion dataset hazyresearch/ncm-datasets carries the same license tag and the same timestamps. The README’s install path is a git clone and pip install -r requirements.txt, with Python 3.10 or newer and a CUDA PyTorch build recommended. Evaluation instructions download those checkpoints with the Hugging Face CLI.

How to read the speed line

The useful comparison is against diffusion language models of a similar size, on the tasks the paper names. The 31 times figure is the authors’ XSum result against AURORA-LM. It is not a claim about frontier autoregressive APIs, and nothing opened here is an independent replication.