---
title: "Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation"
slug: next-concept-modeling-fast-parallel-coarse-to-fine-text-generation
url: https://listedarticles.com/articles/next-concept-modeling-fast-parallel-coarse-to-fine-text-generation
canonical_url: https://arxiv.org/abs/2610.14162
content_type: research
language: en
published_at: 2026-10-09T00:00:00.000Z
updated_at: 2026-11-09T11:14:12.337Z
author: "Hermann Kumbong et al."
authored_by: human
publisher: "arXiv"
publisher_url: https://arxiv.org/
topics: ["Machine Learning", "LLMs", "Research"]
license: arXiv-nonexclusive-distrib-1.0
word_count: 212
reading_minutes: 1
citation: "Hermann Kumbong et al., arXiv. \"Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation.\" 9 Oct 2026. https://arxiv.org/abs/2610.14162 (arXiv-nonexclusive-distrib-1.0)"
---

# Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation

> Introduces Next-Concept Modeling (NCM), a coarse-to-fine alternative to diffusion language models: a multiscale residual VQ-VAE encodes text into hierarchical codes and a transformer predicts each level in parallel, beating or matching the strongest DLM baselines on LM1B, OpenWebText and XSum while running 10-31x faster.

# Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation

**Authors:** Hermann Kumbong, Ishan S. Khare, Iddah Mlauzi, Andrew Woen, Charles Ding, E. Kelly Buchanan, Simran Arora, Azalia Mirhoseini, Christopher Ré

Submitted 9 October 2026 · [arXiv:2610.14162](https://arxiv.org/abs/2610.14162) · [PDF](https://arxiv.org/pdf/2610.14162)

## Abstract

Diffusion language models (DLMs) offer parallel sampling for faster generation than autoregressive models, an increasingly important advantage as inference becomes a dominant computational workload. However, DLMs degrade quickly with fewer sampling steps, process the full sequence at every step, and require many passes over the dataset to reach good quality. We introduce Next-Concept Modeling (NCM), a two-stage, coarse-to-fine framework: a multiscale residual VQ-VAE encodes text into a hierarchy of discrete codes, and a transformer generates one level at a time, predicting all codes within a level in parallel. At matched parameter scales, NCM achieves better generative perplexity than the strongest DLM baseline on LM1B while running $21\times$ faster. On OpenWebText, NCM is competitive with the strongest DLM while running $10\times$ faster; on XSum, it matches the strongest DLM's ROUGE-L while being $31\times$ faster. On the reasoning task of Sudoku, NCM outperforms autoregressive and diffusion baselines across all difficulty levels, using fewer function evaluations. On the reasoning task of Sudoku, NCM outperforms autoregressive and diffusion baselines across all difficulty levels, using fewer function evaluations.
