PaperlayerView exact revision on arXiv

Source-linked AI summary

DiffResearch: Accelerating Deep Research with Diffusion Language Models

Alan-Barsag Gazzaev, Alexey Gavrilov, Mikhail Mozikov, Sergey Muravyov

arXiv:2610.28909v1cs.AI

TL;DR

Autoregressive decoding is sequential, and diffusion models’ latency effects in agentic deep-research pipelines have not been directly measured. DiffResearch compares diffusion and autoregressive backbones in a shared multi-agent pipeline and uses intermediate denoising states for fast decomposition. Diffusion decoding yields a 6× end-to-end speedup over autoregressive decoding without lowering RACE quality, with gains concentrated in planning and long-form writing.

  • Problem

    Autoregressive decoding is sequential, and diffusion models’ latency effects in agentic deep-research pipelines have not been directly measured.

    Source: §1 · pp. 1–2

  • Method

    DiffResearch compares diffusion and autoregressive backbones in a shared multi-agent pipeline and uses intermediate denoising states for fast decomposition.

    Source: §1 · p. 2

  • Results

    Diffusion decoding yields a 6× end-to-end speedup over autoregressive decoding without lowering RACE quality, with gains concentrated in planning and long-form writing.

    Source: §1 · p. 2, §5.6 · p. 6

  • Takeaways & Limitations

    The decoding gain does not resolve retrieval relevance or evidence sufficiency, which remain shared bottlenecks.

    Source: §5.6 · p. 6

    Latency results use an H200 pod and one comparable open-model pair; retrieval was restricted to Semantic Scholar, and the leaderboard comparison is system-level rather than causal evidence about diffusion.

    Source: Limitations · pp. 6–7

Abstract

from arXiv · show

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. We present DiffResearch, an open multi-agent deep research framework that places a diffusion language model (dLLM) at the core of every stage, and measure how a diffusion backbone changes latency and report quality. Holding the agentic scaffold, retrieval code, prompts, and evaluation fixed, we swap the backbone between an open diffusion model and a comparable AR model. Across 100 Deep Research Bench tasks, the diffusion backbone is roughly 6 times faster end-to-end, driven by order-of-magnitude reductions in planning and writing. We further introduce dFast, a decomposition mode that reads intermediate denoising states and extracts subqueries before generation finishes, providing a 1.99 times median speedup for decomposition in isolation with no loss of report quality. On the benchmark's RACE rubric, dFast obtains the best overall score among the three configurations. Finally, in a public system-level leaderboard snapshot, an instance built on a stronger diffusion model ranks above several commercial systems. We release DiffResearch under an Apache-2.0 license.

1 Introduction

Deep research systems rely on sequential autoregressive generation, while diffusion integration into agentic research pipelines and its latency effects have remained underexplored. DiffResearch enables a controlled backbone comparison and reports faster end-to-end performance, alongside a faster decomposition mode and a system-level leaderboard result.

1 Introduction

  • Autoregressive decoding makes planning and long-form writing sequential, while diffusion models refine whole sequences in parallel across denoising steps.Diffusion integration into agentic pipelines and its practical latency consequences had not been directly measured.Source: §1 · pp. 1–2
  • DiffResearch uses diffusion models across its agents and supports a controlled comparison by swapping the backbone while keeping the scaffold fixed.Its fast mode reads usable subqueries from intermediate denoising states before generation completes.Source: §1 · p. 2
  • Diffusion backbones are roughly 6× faster end-to-end on Deep Research Bench, with the gain concentrated in planning and writing.The comparison holds the agentic scaffold, retrieval code, prompts, and evaluation fixed.Source: §1 · p. 2
  • dFast cuts decomposition latency by about half (1.99× median speedup over 100 queries) while preserving downstream report quality, but adds no end-to-end latency reduction over standard diffusion generation.It also obtains the best overall RACE score among the three configurations.Source: §1 · p. 2
  • A public system-level leaderboard snapshot places a stronger-diffusion DiffResearch instance above several listed commercial systems, without isolating diffusion’s effect.The snapshot compares complete systems rather than decoding backbones alone.Source: §1 · p. 2

2 Related Work

Prior work covers deep-research agents, modular agentic scaffolds, evaluation, and diffusion language models, but not a controlled measurement of diffusion decoding in an end-to-end deep-research pipeline. DiffResearch addresses that measurement gap and distinguishes its fast decomposition mode from speculative decoding.

2 Related Work

  • Deep-research systems combine multi-step reasoning, retrieval, and synthesis, while most rely on autoregressive models and sequential decoding.Related work also covers modular planning, retrieval, and evaluation components.Source: §2 · p. 2
  • Diffusion language models refine whole sequences through iterative denoising and can support structure-first generation and iterative error correction.Recent models are described as approaching autoregressive quality while offering speed and global-coherence advantages.Source: §2 · p. 2
  • To the authors’ knowledge, prior work had not measured diffusion decoding’s effects on end-to-end latency and report quality under a controlled backbone swap.The comparison concerns an agentic deep-research pipeline.Source: §2 · p. 3
  • Fast decomposition reads the diffusion model’s own intermediate denoising states and stops when a valid structured output is available, rather than using a draft model.The paper frames this approach as complementary to speculative decoding.Source: §2 · p. 3

3 System Design

DiffResearch uses a modular multi-agent pipeline with interchangeable model backbones and explicit operating modes. Its design preserves shared prompts, retrieval, and evaluation for controlled comparisons, while fast mode extracts valid subqueries from intermediate denoising states.

3.1 Multi-Agent Diffusion Pipeline

  • Nine agents handle query formatting, planning, plan-checking, relevance filtering, information extraction, summarization, complexity assessment, decomposition, and judging.The pipeline has five core stages—intent classification, query reformulation, planning, retrieval, and writing—with judging optional.Source: §3.1 · p. 3

3.2 Retrieval Backends

  • The framework supports web and academic retrieval, but the controlled benchmark runs use only Semantic Scholar, with two results per query.Retrieval uses deterministic API calls without an LLM in the loop; the broader academic backend also includes arXiv.Source: §3.2 · p. 3

3.3 Model Serving and Decomposition

  • The controlled configurations share pipeline code, prompts, retrieval, and evaluation; only the generator backbone and decomposition mode differ.The two backbones are run through Hugging Face runtimes, and both decomposition modes require valid JSON with the requested number of nonempty subqueries.Source: §3.3 · p. 3

3.4 Operating Modes

  • Full mode runs diffusion generation to completion and accepts the final output when it contains the requested number of subqueries.Fast mode instead accepts an exact-count JSON draft from an intermediate denoising state before final generation completes.Source: §3.4 · p. 4

4 Experimental Setup

The evaluation uses Deep Research Bench tasks and RACE report scoring, comparing diffusion and autoregressive backbones under shared pipeline conditions. It also includes isolated decomposition timing and artifact-based diagnostics with a limited manual source-grounded audit.

4 Experimental Setup

  • The benchmark covers 100 PhD-level tasks across 22 fields, split evenly between Chinese and English, and RACE scores reports on comprehensiveness, insight, instruction-following, and readability.RACE uses Gemini 2.5 Pro as evaluator, and scores are reported on a 0–1 scale.Source: §4 · p. 4
  • The controlled comparison runs diffusion and autoregressive backbones of comparable size on an H200 pod, with shared agent code, retrieval, and RACE evaluation.Retrieval is local and restricted to Semantic Scholar, with four subqueries, two results per query, and one request per second.Source: §4 · p. 4
  • Both diffusion modes are about 6× faster than AR across the 100 benchmark tasks, while dFast adds no end-to-end reduction over dFull.The cited table reports end-to-end latency in seconds.Source: Table 1 · p. 4
  • A separate 100-query study measures decomposition-stage latency through the diffusion streamer, excluding model loading.Saved reports are also analyzed for length, citations, headings, and language-specific RACE scores; a blinded audit checks five claims and all citation mentions in 10 reports.Source: §4 · p. 4

5 Results and Discussion

Across 100 Deep Research Bench tasks, diffusion configurations were about six times faster end-to-end than AR, with gains concentrated in planning and writing. dFast accelerated decomposition in isolation; quality results and a separate leaderboard snapshot provide additional context, with the latter not isolating diffusion’s effect.

5.1 End-to-End Latency

  • AR took 71.45 s mean end-to-end, versus 11.89 s for dFull and 11.97 s for dFast across 100 tasks; dFast did not improve on dFull.The p90 was 111.87 s for AR versus 16.13 s for dFull and 16.33 s for dFast.Source: §5.1 · p. 4

5.2 Where the Time Goes

  • Planning fell from 39.32 s to about 3.6–3.7 s (10.7×), while article writing fell from 59.78 s to about 5.1–5.3 s (11.7×).Retrieval was essentially unchanged across backbones because it uses a shared deterministic Semantic Scholar call; it was slightly slower in diffusion runs as more tasks reached retrieval.Source: §5.2 · p. 5, Table 2 · p. 5

5.3 Accelerating Decomposition

  • 1.99× median speedup came from taking the first exact-count draft rather than waiting for full generation in a 100-query decomposition probe.A stable object was only 1.04× faster at the median, but matched the final answer in 91 of 100 cases, versus 1 of 100 for the first draft.Source: §5.3 · p. 5, Table 3 · p. 5
  • Because subqueries feed retrieval that tolerates wording noise, dFast uses the first exact-count draft for decomposition, not for final writing.The decomposition improvement is stage-level and adds no end-to-end reduction in the current pipeline.Source: §5.3 · p. 5

5.4 Report Quality

  • dFast had the highest available-case overall RACE score (0.3969), versus 0.3895 for dFull and 0.3904 for AR; paired common-success intervals included zero in both languages.Diffusion reports were about 10–17% shorter than AR reports, while the writer was about 11–12× faster; shorter outputs alone therefore cannot explain the speedup.Source: §5.4 · pp. 5–6
  • In a blinded audit of 10 reports, 26 of 50 sampled material claims (52%) were not established by the stored source packet, identifying retrieval grounding as a shared bottleneck.The per-mode audit samples were too small for comparative claims.Source: §5.4 · p. 6

5.5 System-Level Leaderboard Snapshot

  • The Mercury-backed DiffResearch system ranked above several listed commercial systems in a public leaderboard snapshot, but the comparison does not isolate diffusion decoding.The snapshot’s RACE overall score was 48.03, versus 46.45 for OpenAI Deep Research, 45.49 for Dr. Tulu, and 45.00 for Claude Research.Source: Table 5 · p. 6, §5.5 · p. 6

5.6 Discussion

  • With the backbone as the only change, diffusion made planning and long-form writing about an order of magnitude faster, yielding a 6× end-to-end speedup without lowering RACE quality.The leaderboard supports system-level competitiveness, not attribution of the ranking to diffusion decoding.Source: §5.6 · p. 6

Limitations

The reported latency results are bounded by the tested hardware, retrieval setup, and model pair. The study also limits fast mode to decomposition and identifies constraints on generalizing its system-level comparisons.

Limitations

  • Fast mode is scoped to decomposition and did not reduce observed end-to-end latency relative to dFull in the current setup.Its first exact-count draft matched the final string in only 1 of 100 probes, so it is intended for downstream consumers that tolerate wording noise, not final answers.Source: Limitations · pp. 6–7
  • Latency gaps may differ when local diffusion inference is the bottleneck or retrieval costs little relative to generation, and tight consumer-grade budgets were not benchmarked.Source: Limitations · p. 6
  • The backbone comparison uses one pair of comparable-size open models, so a broader sweep across model families and sizes would strengthen attribution.Source: Limitations · p. 7
  • The leaderboard result compares full systems in a time-specific snapshot, not writer-isolated variants, so it does not attribute performance causally to diffusion decoding alone.Source: Limitations · p. 7
  • Retrieval used Semantic Scholar only; some tasks lacked needed sources, while arXiv and web retrieval were disabled.Source: Limitations · p. 7
  • Parallel retrieval dispatch was not evaluated because requests were rate-limited to one per second; testing it with a higher-throughput backend remains future work.Source: Limitations · p. 7

Related papers

Loading 2610.28909v1…