Dev reproduces Google query fanout paper: diffusion model outputs vectors, skipping text

dejanseo · x · 2026-09-24

dejanseo reproduced Google's new query fanout research, trained the model himself, and argues that if it ships, text fanouts are gone.

Key insight: the research skips text generation entirely—a small diffusion model is trained on vectors and outputs vectors, enabling direct semantic document retrieval.
He simplified the pipeline by skipping RL, using teacher models to generate quality synthetic fanout data.
The trained model is open-sourced on Hugging Face (dejanseo/fanout-diffusion) with a live Space demo that uses embeddings instead of text queries, trained on Google's e-commerce taxonomy fanouts.
He reports blazing speeds using a 1-bit architecture with custom CUDA kernels.

More from Infra