Twelve Steps, Coarse to Fine

Diffusion language models promise parallel generation, then spend 64 to 1,024 passes over the whole sequence to deliver it. Next-Concept Modeling, NCM, is a new open-source answer from the Hazy Research group, released with a paper, arXiv:2610.14162, submitted on 9 October. Its pitch is that a text sample should be written the way a sketch is drawn: a rough version first, then finer ones. On LM1B it uses 12 sampling steps where the diffusion baselines use 128.

I read the paper and the README instead of the thread. The speed numbers hold up on the page. What they are measured against needs a closer look.

What a concept is

NCM has two stages. First a multiscale residual VQ-VAE turns a sequence into a hierarchy of discrete codes. Level one is a very short code sequence, and each later level is longer and encodes what the earlier levels missed, until the last level reconstructs the tokens. In the paper’s ablation the ladders start at one position and end at 128. Second, a transformer is trained to produce these levels in order. Within a level, every position is predicted in parallel, so a level costs one forward evaluation, and the coarse levels are cheap because they are short.

That last property explains the speed. A diffusion model that takes 128 steps runs each step over the full sequence. On LM1B NCM takes about a tenth as many steps, and the early ones run over a fraction of the sequence. The paper’s tokenizer is close to lossless where it needs to be: exact token reconstruction is 99.86 percent on LM1B, 99.99 on OpenWebText, 98.79 on WMT14 and 100 percent on Sudoku, where one wrong digit voids a solution.

Where the speedup comes from

The paper times 1,024 sequences at batch size 16 on a single NVIDIA H100 and compares against the strongest diffusion baseline on each task. On LM1B that baseline is DUO. NCM uses 11 times fewer function evaluations and runs 21 times faster. On XSum the baseline is AURORA-LM, and NCM uses 16 times fewer evaluations and runs 31 times faster. The wall-clock ratio is larger than the step ratio in both cases, which is the short-coarse-levels effect: the average step is also cheaper.

NCM's reduction in sampling steps versus its wall-clock speedup, on LM1B and XSum A dumbbell chart on a 0 to 40 scale. On LM1B against DUO, NCM uses 11 times fewer function evaluations and runs 21 times faster. On XSum against AURORA-LM, NCM uses 16 times fewer function evaluations and runs 31 times faster. In both rows the speedup exceeds the reduction in steps. NCM VS STRONGEST DIFFUSION BASELINE · H100, 1,024 SEQUENCES AT BATCH 16 · ARXIV 2610.14162 0x 10x 20x 30x 40x LM1B, vs DUO XSum, vs AURORA-LM 11x fewer function evaluations on LM1B — arXiv 2610.14162, section 4.6 16x fewer function evaluations on XSum — arXiv 2610.14162, section 4.6 21x faster wall-clock on LM1B — arXiv 2610.14162, section 4.6 31x faster wall-clock on XSum — arXiv 2610.14162, section 4.6 11x21x 16x31x Open circle: fewer function evaluations. Filled: faster wall-clock. Both as reported by the paper.
Cutting steps is only part of it. Each NCM step is also cheaper than a diffusion step, so wall-clock gains run well above the step ratio.

One number does not line up between the sources. The README says 9 to 31 times faster, while the paper’s abstract and contributions list say 10 to 31. The OpenWebText speedup is the low end in the paper at 10 times. The README gives no per-task breakdown, and I could not find a 9 in the paper. I use the paper’s figure.

What “competitive” covers

The quality claims differ by row.

Unconditional text. Quality is generative perplexity under GPT-2 large, paired with unigram entropy so that degenerate samples show up. On LM1B, NCM scores 93.05 at 12 steps against 97.30 for DUO at 128, with entropy 4.31 and 4.30, and the real data scores 53.51. That is a win, and a modest one. On OpenWebText it is 83.12 at 32 steps against 77.69 for DUO at 1,024, so NCM is behind on this metric and the paper’s word for it, “competitive,” is accurate. Ground-truth text scores 14.15 on OpenWebText, and an autoregressive sampler scores 35.45, so every diffusion-family model in the table, NCM included, is far from the data on this metric. The paper’s own table also shows why the metric needs the entropy column: greedy autoregressive decoding scores 1.07 with entropy 1.65, which is repetition, not quality.

Summarization. Without guidance, NCM gets ROUGE-1/2/L of 35.0, 12.8 and 28.0 at 8 steps, against AURORA-LM’s 34.7, 12.2 and 26.6 at 128. That is the cleanest win in the paper, with a smaller 88M generator against a 130M one. With minimum Bayes risk decoding, where the saved time pays for 128 evaluations of candidate samples, it reaches 37.3, 14.0 and 29.6, ahead of the guided AURORA-LM row at 184 evaluations.

Translation. This is the soft spot. On WMT14 German to English NCM scores 22.6 BLEU at 11 steps. Among unguided non-autoregressive models only E2D2 beats it, at 24.8 and 64 steps, and the autoregressive baseline scores 25.2. NCM with MBR matches that 25.2, but a guided ELF row scores 26.4.

Sudoku. NCM solves 99.5, 96.5 and 86.1 percent of easy, medium and hard boards with 12 steps. The strongest diffusion baseline, FMLM+, solves 97.7, 92.8 and 81.3 with 16, and DUO scores 58.4 on hard. A small autoregressive model solves 0.6 percent of hard boards with sampling and 1.0 percent with greedy decoding. The paper’s explanation fits the design: later levels can correct earlier partial solutions, which left-to-right decoding cannot do.

Two caveats apply across all of this. Many of the baseline numbers for conditional tasks are taken from earlier papers, marked in the table, so the comparison is against published settings, not a rerun under one harness. And the NCM generators here are small, roughly 100 million parameters, with a scaling study up to 725 million on LM1B. The paper shows nothing at the sizes where language models are usually deployed.

Cheaper to train, and what is still fixed

The paper also reports training on about 4.9 times fewer token positions than the diffusion baselines on LM1B, 13.36 billion against 66 billion, counting tokenizer and generator together. On OpenWebText the total is about 127 billion against more than 500 billion for MDLM and DUO. The authors note these are token-exposure counts, not FLOPs.

The limitations section names the design costs. The hierarchy depth and per-level lengths are fixed, so even a short continuation walks the whole ladder. The tokenizer is trained separately from the generator. Samples are “generally fluent” with occasional artifacts at coarse-level transitions.

My read: the idea I expect to travel is the cost structure, not the benchmark wins. A generator whose early steps are short is cheap in a way that more careful diffusion samplers are not, and the spare time can be spent on candidates and selection, as the MBR rows show. Whether it survives scale is untested. The repository is MIT licensed and ships checkpoints for XSum, WMT14, Sudoku, LM1B, TinyStories and OpenWebText, with one evaluation script, evaluate.py, and a config per task. The README links the checkpoints on Hugging Face, so the claims can be rerun.

References

  1. Hermann Kumbong, Ishan S. Khare, Iddah Mlauzi, Andrew Woen, Charles Ding, E. Kelly Buchanan, Simran Arora, Azalia Mirhoseini, Christopher Ré, “Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation”, arXiv:2610.14162, submitted 9 October 2026.
  2. Hazy Research, HazyResearch/ncm README, accessed 20 October 2026.