Stanford's Hazy Research Proposes Next-Concept Modeling as Alternative to Token-by-Token AI Generation

Researchers at Stanford University's Hazy Research group have introduced Next-Concept Modeling, a new language modeling architecture released in October 2026. Unlike traditional autoregressive LLMs, which generate text one token at a time, the approach first builds a global representation of the response and then refines it in parallel. It uses a multiscale hierarchy of discrete codes, predicting all codes at each level simultaneously and conditioning each subsequent level on prior outputs. The reported results include 10 to 31 times faster inference than evaluated diffusion baselines and roughly 4 to 5 times fewer training tokens than certain baselines. The researchers note these results do not establish performance advantages over widely used models such as GPT, Qwen, or Llama.