DiffResearch puts diffusion models at the core of deep research agents
Original title: DiffResearch: Accelerating Deep Research with Diffusion Language Models
On 100 benchmark tasks the authors report about a 6x end-to-end speedup versus a comparable autoregressive model.
Researchers introduce DiffResearch, an open multi-agent research framework that uses a diffusion language model for planning, evidence gathering, synthesis, and writing. With the agent scaffold, retrieval, prompts, and evaluation held fixed, they compare an open diffusion backbone with a comparable autoregressive model. Across 100 Deep Research Bench tasks, they report roughly six times faster end-to-end runs, mainly from faster planning and writing. A mode called dFast reads intermediate denoising states to emit subqueries early; the authors say it scores best on the RACE rubric among three setups. The code is released under Apache-2.0.
Unconfirmed: Latency, quality, and leaderboard claims are the authors' results and have not been independently confirmed.
- dFast is reported to give a 1.99 times median speedup for decomposition alone.
- Authors say dFast did not reduce report quality.
- A stronger diffusion instance is said to rank above several commercial systems on a public snapshot.
- The framework is released under the Apache-2.0 license.
From the original article
1 min read at huggingface.co
Deep research systems iteratively plan, gather evidence, synthesize, and revise, making end-to-end latency a key bottleneck for interactive use. Almost all production systems are built on autoregressive (AR) language models and inherit left-to-right decoding in planning and long-form writing.
By Alan-Barsag Gazzaev, Alexey Gavrilov, Mikhail Mozikov · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.