Stanford Team Publishes Next-Concept Modeling Paper on Faster Text Generation

A team of researchers working in the area Perplexity AI draws its own models and infrastructure thinking from has published a paper describing a new way to generate text that sidesteps a weakness of diffusion language models. The paper, posted to arXiv on October 9, 2026, by Hermann Kumbong and eight co-authors including Simran Arora, Azalia Mirhoseini and Christopher Re, introduces Next-Concept Modeling, or NCM.

Diffusion language models have become attractive because they sample many tokens at once rather than one at a time, which should make them faster than the autoregressive models that power most chat products today, including the one behind an answer engine like Perplexity. The paper’s authors note that this promise breaks down in practice: diffusion models lose quality quickly once the number of sampling steps is reduced, they have to process an entire sequence at every single step regardless of how little of it is uncertain, and they need many passes over training data before reaching usable quality.

NCM’s answer is a two stage pipeline. First, a multiscale residual VQ-VAE turns a passage of text into a hierarchy of discrete codes, coarse structure at the top, fine detail further down. Second, a transformer generates that hierarchy one level at a time, predicting every code inside a level in parallel rather than working through the sequence token by token. The idea borrows the logic of coarse-to-fine image generation and applies it to language.

The numbers

Across three benchmarks the authors report sizable speed gains alongside matched or better quality. On the LM1B benchmark, NCM beat the strongest diffusion baseline on generative perplexity while running faster. On OpenWebText it stayed competitive with the strongest diffusion model while running faster still. On the summarization benchmark XSum, it matched the strongest diffusion baseline’s ROUGE-L score while running faster than that too. On the Sudoku reasoning task, the authors report that NCM outperformed both autoregressive and diffusion baselines across every difficulty level tested, while using fewer function evaluations per solution.

For a company whose entire product is an answer engine rather than a model, work like this is a reminder that the inference costs behind a query, not just the headline model behind it, decide what a subscription can afford to serve. If generation methods like NCM close the quality gap between fast parallel sampling and slower step-by-step decoding, a provider could serve more detailed answers at the same compute budget, or hold prices steady while demand on inference keeps climbing. The paper does not describe any product deployment, so the practical change for Perplexity’s users, if any, is still a few steps away.