Next-Concept Modeling claims 31× speedup over diffusion LMs

A new method called Next-Concept Modeling claims to beat the strongest diffusion language model on generative perplexity while running 21× faster at matched parameter scales.

The paper, posted on arXiv by Hermann Kumbong, Simran Arora, Azalia Mirhoseini, Christopher Ré and five co-authors, works in two stages: a multiscale residual VQ-VAE encodes text into a hierarchy of discrete codes, then a transformer 'generates one level at a time, predicting all codes within a level in parallel.' Two of the researchers we follow had already circulated it.

The argument is pointed straight at diffusion. 'Diffusion language models (DLMs) offer parallel sampling for faster generation than autoregressive models, an increasingly important advantage as inference becomes a dominant computational workload,' the abstract says, before listing the camp's weaknesses: 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.'

On OpenWebText the paper reports NCM is 'competitive with the strongest DLM while running 10× faster'; on XSum, it 'matches the strongest DLM's ROUGE-L while being 31× faster.' The abstract publishes no per-task perplexity figures, no model-scale numbers, and no comparison to autoregressive serving tricks like speculative decoding.