Diffusion and Continuous Generative Models Challenge Autoregressive Defaults
Key Questions
What is DiffusionGemma and who announced it?
DiffusionGemma is a 26B parameter text-diffusion model announced by Raia Hadsell at RAAIS 2026. It can self-correct during generation, challenging autoregressive approaches.
How does DiffusionGemma differ from traditional language models?
It uses a diffusion process rather than autoregressive generation, enabling mid-generation self-correction. This offers potential gains in reasoning and controllability.
What other model was revealed alongside DiffusionGemma?
Genie-3 was also revealed, demonstrated on Street View data. It extends generative capabilities in visual domains.
DiffusionGemma, UnMaskFork, Hierarchical Continuous Diffusion Language Models, Multimodal Flow, and adaptive reward routing show growing momentum for self-correcting, searchable, continuous, and multimodal generation. Structured-task and audio-video results are encouraging, but evidence at larger scale and on reasoning, controllability, efficiency, and reproducibility remains limited.