AI Trend Research
EN

← AI glossary

AI glossary · Basic · Language models · Efficiency & infra

Next-Concept Modeling

Next-Concept Modeling (NCM) is a text generation method that does not write text one word at a time. It first builds a rough outline of the whole text as coarse codes, then fills in the details on top of it, all at once at each stage, to produce text quickly. It was proposed in a research paper posted on arXiv in October 2026.

About this term
Level
Basic · Technical terms you often meet in AI news
Added
2026-10-24
In plain words It is similar to drawing a picture: you first sketch the overall composition in pencil, then paint in the large areas of color, and finally refine the details. The analogy breaks down in one respect, though: the "concepts" here are not concept words that people use, but groups of codes the model learned on its own.

What is different

Most LLMs in use today are autoregressive models, so they look at the previous tokens and append the next token one at a time. This is accurate, but generation time grows as the text gets longer. Diffusion language models (DLMs) produce multiple tokens at the same time and are faster, but according to the paper, quality drops quickly when the number of steps is reduced, the entire sequence is reprocessed at every step, and they need multiple passes over the dataset during training to reach good quality. NCM tries to avoid these problems with a two-stage structure that goes from coarse to fine (coarse-to-fine).

Three ways to generate text

Autoregressive modelOne token at a time, in order

  • Predicts the next token from previous tokens
  • Generation time grows with text length

Diffusion language modelMultiple tokens at the same time

  • Faster through parallel generation
  • Quality drops with fewer steps
  • Processes the full sequence every step

NCMOutline first, details later

  • Generates level-by-level codes from the top level
  • Predicts each level all at once in parallel
  • Still at the research stage

공통 All three generate text with transformer-based models

A comparison of differences in generation order and parallelism. For performance figures, see the researchers' claims in the main text.

How it works

In the first stage, a compression model called a VQ-VAE converts text into indices from a fixed list of codes. It splits this into multiple levels (multiscale): the upper levels capture the broad outline of the text, and the lower levels fill in the differences the upper levels missed (residuals). In the second stage, a transformer generates these codes one level at a time, starting from the top level. Codes within the same level are predicted all at once in parallel rather than in order. Once all levels are filled, the codes are decoded back into text.

NCM's coarse-to-fine generation flowOriginal text → VQ-VAE encoder; VQ-VAE encoder → Level-by-level codes; Level-by-level codes → Transformer(Training); Transformer → Upper-level codes; Upper-level codes → Lower-level codes(One level at a time); Lower-level codes → Decoder; Decoder → Generated text① Compress text into level-by-level codes② Generate from the top levelOriginal textVQ-VAE encoderMultiscale · residualLevel-by-levelcodesOutline → detailsTransformerUpper-level codesBroad outline,parallel predictionLower-level codesDetail differences,parallel predictionDecoderGenerated textTrainingOne level at a time
NCM's coarse-to-fine generation flow The top row is the training process that compresses text into level-by-level codes, and the bottom row is the generation process in which a transformer fills in codes in parallel one level at a time from the top level, then turns them back into text.

Why it is getting attention

As inference has come to account for a large share of AI computation, ways to produce text of the same quality faster have become important. The researchers say that, comparing models of similar size, NCM scored better on a generation quality metric than the strongest DLM on the LM1B dataset while being 21 times faster. On OpenWebText, they report quality similar to DLMs at 10 times the speed, and on the summarization task XSum, the same level of score 31 times faster. They also say that on reasoning problems such as Sudoku, it solved problems better at every difficulty level with less computation than autoregressive and diffusion methods. General outlets such as Forbes have also described it as a method that generates "concepts first, words later."

Things to keep in mind

As of October 2026, NCM is at the stage of a research paper posted on arXiv, and no commercial service using it has been confirmed. The performance figures above are results reported by the researchers; they have not yet been reproduced by other researchers or verified at large model scale. The comparisons were also mainly against diffusion language models, so whether it can replace the large autoregressive LLMs in use today needs to be examined separately. Also, despite the name "concept," what the model produces is not human-readable thought but codes learned by a compression model.

Example

There are no cases yet of NCM being used in an actual product. One scenario you can imagine is a service that summarizes a long article in one go. An autoregressive model has to write the summary one token at a time from the beginning, but with the NCM approach, it could first set the outline of the whole summary and then fill in details all at once at each level, returning an answer faster. The paper's reported speed advantage on a summarization task (XSum) is an experiment with this kind of use in mind.

Related terms

Autoregressive model Diffusion model Transformer Token Inference Speculative decoding

References