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

Researchers led by Hermann Kumbong introduced Next-Concept Modeling (NCM), a two-stage coarse-to-fine text generation framework that pairs a multiscale residual VQ-VAE with a transformer that predicts all codes within a level in parallel. At matched parameter scales, NCM beat the strongest diffusion language model baseline on generative perplexity on LM1B while running 21 times faster, and on OpenWebText it was competitive with that baseline at 10 times the speed. On XSum, NCM matched the top diffusion model's ROUGE-L score while running 31 times faster, and on Sudoku reasoning it outperformed both autoregressive and diffusion baselines across all difficulty levels with fewer function evaluations. The results matter because inference cost has become a dominant computational workload, and NCM offers a path to faster generation without sacrificing quality at comparable model sizes. Next to watch: whether the approach scales to larger models and broader benchmarks, and whether the efficiency gains hold outside the tasks reported in the preprint.