DiffResearch framework uses diffusion models to cut research AI latency

DiffResearch is a new multi‑agent research framework that puts a diffusion language model (dLLM) at the core of planning, retrieval and writing. The authors compared the diffusion backbone with a similar autoregressive model on 100 Deep Research Bench tasks. The diffusion version ran roughly six times faster overall, mainly because planning and long‑form writing became much quicker. A special mode called dFast reads intermediate denoising states and splits the work into subqueries, giving an additional median 1.99× speed boost without hurting report quality. The code is released under an Apache‑2.0 license.