A coarse-to-fine way to generate text: a model predicts a hierarchy of codes from broad structure down to detail, in parallel rather than token by token.
Last verified · 2026-11-09 · by Moe Ameen
Next-Concept Modeling (NCM) is a text-generation approach that builds a passage coarse-to-fine — predicting a broad outline of the output first, then filling in progressively finer detail — instead of writing one word after another. It was introduced in the October 2026 paper "Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation" from HazyResearch, Christopher Ré's lab at Stanford. The pitch is speed with quality held: by sketching structure before surface text and decoding many positions at once, NCM reaches comparable output quality to a [diffusion language model](/glossary/diffusion-language-model) while running up to roughly 30× faster on the tasks the authors tested.
Mechanically it is two stages. First a tokenizer — a multiscale residual VQ-VAE — encodes text not into a flat list of tokens but into a hierarchy of discrete codes, coarse levels carrying global structure and finer levels adding detail on top of them. Then a transformer generates that hierarchy one level at a time, predicting all of the codes within a level in a single parallel pass and conditioning each new level on the reconstruction of the coarser ones above it. The coarse early passes run on short sequences, so the model needs far fewer sequential forward passes than a method that re-processes the whole sequence at every step — which is where the speedup comes from.
The word "concept" is doing analogy work, and it is worth being precise about it. NCM does not define a concept as a named semantic idea; a "concept" here is a coarse level of that learned code hierarchy — the broad-strokes representation the model commits to before it writes the detail. The name echoes "next-token prediction," swapping the unit from one token to one level of structure, and the picture to hold is an editor who blocks out the shape of a whole answer and then sharpens it in passes, versus a typist producing final words left to right. It is the text cousin of coarse-to-fine, next-scale generation already used in AI image models.
The lineage runs through two research threads converging in 2026. One is the push to make language generation parallel: autoregressive models like GPT and Claude emit one token at a time, and [diffusion language models](/glossary/diffusion-language-model) answered that bottleneck by denoising many positions per step — but they still re-process the full sequence at every step, degrade when you cut the step count, and need many passes over the data to train well. The other thread is coarse-to-fine generation from the image world, where next-scale prediction (generating an image as a stack of progressively finer resolutions) had shown that committing to global structure first, then refining, can be both fast and high quality.
Next-Concept Modeling, posted to arXiv on October 9, 2026 (2610.14162) by a nine-author Stanford team including Hermann Kumbong, Simran Arora, Azalia Mirhoseini, and Christopher Ré, carried that coarse-to-fine idea into text. The authors reported matching or beating the strongest diffusion-LM baselines on language modeling and summarization while running an order of magnitude or more faster — about 21× on LM1B with better generative perplexity, roughly 10× on OpenWebText at competitive quality, and about 31× on the XSum summarization benchmark at matched ROUGE-L — and training with several times fewer tokens. It is early-stage research, released at small, academic model scales with code (MIT-licensed, on GitHub as HazyResearch/ncm) and weights published, not a product you can call from an API. Its real significance is as a third architectural option in a conversation that through 2026 had narrowed to "autoregressive versus diffusion."
| Platform | Behavior |
|---|---|
| vs. autoregressive models (GPT, Claude) | Autoregressive models generate strictly left to right, one token conditioned on all the ones before it, so wall-clock speed is bounded by sequence length. NCM commits to coarse structure first and fills levels in parallel, needing far fewer sequential passes. The trade is maturity: autoregressive models hold the general-quality frontier in 2026, while NCM is a small-scale research result. |
| vs. diffusion language models | Both decode many positions at once. A diffusion LM denoises the full sequence repeatedly and loses quality when you reduce steps; NCM runs its early passes on short coarse sequences and climbs the hierarchy, which is where its reported 10–31× speedups over diffusion baselines come from. NCM is the paper's direct point of comparison. |
| vs. next-scale image generation | NCM is the text analogue of coarse-to-fine, next-scale image models: predict a global sketch, then refine. The shared idea is a learned multi-level code hierarchy generated broad-to-fine; the difference is medium — discrete text codes rather than image scales. |
| vs. ConceptLM / "next concept prediction" | Easy to confuse by name. ConceptLM (a separate 2026 paper) adds a next-concept-prediction objective to an autoregressive model, quantizing groups of tokens into concepts predicted alongside next-token prediction. It is still left-to-right; NCM is a parallel, coarse-to-fine generator. Same vocabulary, different architecture. |
What makes Next-Concept Modeling interesting is not the speedup number, eye-catching as it is — it is the prior baked into the architecture: decide the shape of the whole thing before you commit to the words. That is a genuinely good instinct, and it happens to be exactly how competent content gets made. You do not write a good post or a good video script by guessing word 1, then word 2; you fix the angle and the structure first, then fill in the language. NCM is a small, early research result and I would not bet a workflow on it yet, but the idea it encodes — coarse structure, then fine detail — is the right one, and it is worth watching whether it scales.
The practical translation for anyone making content rather than training models is the same lesson, one level up. The bottleneck in a content operation was never how fast a model emits tokens; it is holding one structure — the angle, the voice, the format — consistent across everything a brand puts out. That "concept before detail" layer is deliberately what a tool like [Kompozy](/) owns: a [Persona Brief](/glossary/persona-brief) fixes the voice and the claims before a single frame or sentence is generated, and the format system carries that structure through net-new formats — persona-led avatar video, brand-exact image templates, carousels — as well as through turning one source into a blog article, a newsletter, and the social spread, so the through-line never drifts regardless of which output it lands in. Kompozy is model-agnostic by design — it runs on Claude and OpenAI today and could adopt a faster decoder the day one is the better fit — because what reaches the person publishing is the finished, on-brand calendar, not the decode speed underneath it. Watch the architecture; build on the operation.
Next-Concept Modeling (NCM) is a coarse-to-fine text-generation method from Stanford's HazyResearch lab, introduced in October 2026. Instead of writing one token at a time, it encodes text into a hierarchy of discrete codes — coarse levels for global structure, finer levels for detail — and a transformer generates one level at a time, predicting all codes in a level in parallel. The result is text produced broad-to-fine, which the authors report is up to roughly 30× faster than a diffusion language model at comparable quality.
It is a coarse level of a learned discrete-code hierarchy, not a named semantic idea. NCM's tokenizer (a multiscale residual VQ-VAE) represents text at several levels of detail; the coarse levels carry global structure and are the broad "concept" the model commits to before filling in finer detail. The name echoes "next-token prediction," swapping the unit from one token to one level of structure.
Both decode many positions at once rather than strictly left to right. A diffusion language model denoises the entire sequence repeatedly and loses quality when you reduce the number of steps. NCM instead generates a hierarchy coarse-to-fine, running its early passes on short coarse sequences, so it needs far fewer sequential forward passes — which is where its reported 10× to 31× speedups over diffusion baselines come from. Diffusion LMs are NCM's main point of comparison in the paper.
It is designed to be faster than parallel diffusion baselines, and by extension faster than one-token-at-a-time autoregressive models like GPT and Claude on the tasks tested. But the reported figures (about 21× on LM1B, 10× on OpenWebText, 31× on XSum) are from small research-scale benchmarks, some against randomly initialized baselines, not production systems. As of its release the strongest general-quality models are still autoregressive; NCM is a promising research direction, not a shipped product.
For content work, practically no. NCM is academic research released with MIT-licensed code (GitHub: HazyResearch/ncm) and published weights at small model scales — reproducible by researchers, not exposed as a hosted API or a consumer tool. If you make content, the architecture that drafts your text matters far less than the system that turns a draft into finished, on-brand posts across platforms, which is a separate, model-agnostic layer.