DiffResearch: Accelerating Deep Research with Diffusion Language Models

DiffResearch. Accelerating Deep Research with Diffusion Language Models. By Alan-Barsag Gazzaev and three others, from ITMO University and AXXX. Published on the twentieth of October. Its code repository is public and real: fourteen source files.

Deep research systems plan, gather evidence, synthesize, and revise, over and over. That loop makes end-to-end latency a key bottleneck whenever someone is waiting on the answer. Almost all production systems are built on autoregressive language models, and so they inherit left-to-right decoding, both in planning and in long-form writing.

DiffResearch is an open multi-agent deep research framework that puts a diffusion language model at the core of every stage. It is also an experiment: the authors use it to measure how a diffusion backbone changes both latency and report quality.

The comparison is kept clean. The agentic scaffold, the retrieval code, the prompts, and the evaluation are all held fixed, and only the backbone is swapped, between an open diffusion model and a comparable autoregressive one. On top of that sits a second contribution, dFast. It is a decomposition mode that reads the intermediate denoising states and pulls subqueries out of them before generation has finished, so the decomposition step does not have to wait for a completed draft. It is judged two ways: on the speed of decomposition by itself, and on the quality of the reports that come out at the end.

Across one hundred Deep Research Bench tasks, the diffusion backbone runs roughly six times faster end-to-end. The abstract attributes that to order-of-magnitude reductions in planning and in writing, the same two stages where autoregressive systems inherit left-to-right decoding.

Measured in isolation, dFast gives a median speedup of one point nine nine times for decomposition, with no loss of report quality. On the benchmark's RACE rubric, it obtains the best overall score of the three configurations. And in a snapshot of a public system-level leaderboard, an instance built on a stronger diffusion model ranks above several commercial systems. DiffResearch is released under an Apache-2.0 license.

The study is built as a controlled swap, and that is what lets the speed be attributed to the backbone rather than to anything around it. Two results sit together. Swapping the backbone and nothing else makes the whole pipeline several times faster. And the diffusion model's intermediate states turn out to be useful before they are finished, since dFast reads subqueries out of them early and loses no quality in doing so.