NCM Records 21x Speedup and Lower Perplexity Than DLMs on LM1B

The arXiv paper 2610.14162 introduces a two-stage coarse-to-fine architecture. A multiscale residual VQ-VAE first compresses text into hierarchical discrete codes. A transformer then predicts each level in parallel rather than token by token. This replaces the full-sequence denoising passes required by diffusion language models and reduces the number of function evaluations needed for high-quality output. Benchmarks show clear gains. On LM1B, NCM improves perplexity over the leading DLM while delivering a 21x wall-clock reduction. On OpenWebText it matches DLM quality at 10x speed. On XSum it equals the strongest DLM ROUGE-L score at 31x speed. Sudoku experiments further demonstrate fewer evaluations across all difficulty tiers compared with both autoregressive and diffusion baselines. The design directly attacks the inference bottleneck that has grown dominant as sequence lengths increase. Parallel level generation exploits the hierarchy produced by the VQ-VAE, an approach absent from prior diffusion or autoregressive work. Production systems that currently amortize long diffusion schedules stand to cut latency and energy per token without quality loss if the VQ-VAE hierarchy scales to larger vocabularies.