Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Marco DeepResearch: Unlocking Efficient Deep Research Agents via Verification-Centric Design

Bin Zhu, Qianghuai Jia, Tian Lan, Junyang Ren, Feng Gu, Feihu Jiang, Longyue Wang, Zhao Xu, Weihua Luo

Abstract

Deep research agents autonomously conduct open-ended investigations, integrating complex information retrieval with multi-step reasoning across diverse sources to solve real-world problems. To sustain this capability on long-horizon tasks, reliable verification is critical during both training and inference. A major bottleneck in existing paradigms stems from the lack of explicit verification mechanisms in QA data synthesis, trajectory construction, and test-time scaling. Errors introduced at each stage propagate downstream and degrade the overall agent performance. To address this, we present Marco DeepResearch, a deep research agent optimized with a verification-centric framework design at three levels: \textbf{(1)~QA Data Synthesis:} We introduce verification mechanisms to graph-based and agent-based QA synthesis to control question difficulty while ensuring answers are unique and correct; \textbf{(2)~Trajectory Construction:} We design a verification-driven trajectory synthesis method that injects explicit verification patterns into training trajectories; and \textbf{(3)~Test-time scaling:} We use Marco DeepResearch itself as a verifier at inference time and effectively improve performance on challenging questions. Extensive experimental results demonstrate that our proposed Marco DeepResearch agent significantly outperforms 8B-scale deep research agents on most challenging benchmarks, such as BrowseComp and BrowseComp-ZH. Crucially, under a maximum budget of 600 tool calls, Marco DeepResearch even surpasses or approaches several 30B-scale agents, like Tongyi DeepResearch-30B.

Marco DeepResearch: Unlocking Efficient Deep Research Agents via Verification-Centric Design

Abstract

Deep research agents autonomously conduct open-ended investigations, integrating complex information retrieval with multi-step reasoning across diverse sources to solve real-world problems. To sustain this capability on long-horizon tasks, reliable verification is critical during both training and inference. A major bottleneck in existing paradigms stems from the lack of explicit verification mechanisms in QA data synthesis, trajectory construction, and test-time scaling. Errors introduced at each stage propagate downstream and degrade the overall agent performance. To address this, we present Marco DeepResearch, a deep research agent optimized with a verification-centric framework design at three levels: \textbf{(1)~QA Data Synthesis:} We introduce verification mechanisms to graph-based and agent-based QA synthesis to control question difficulty while ensuring answers are unique and correct; \textbf{(2)~Trajectory Construction:} We design a verification-driven trajectory synthesis method that injects explicit verification patterns into training trajectories; and \textbf{(3)~Test-time scaling:} We use Marco DeepResearch itself as a verifier at inference time and effectively improve performance on challenging questions. Extensive experimental results demonstrate that our proposed Marco DeepResearch agent significantly outperforms 8B-scale deep research agents on most challenging benchmarks, such as BrowseComp and BrowseComp-ZH. Crucially, under a maximum budget of 600 tool calls, Marco DeepResearch even surpasses or approaches several 30B-scale agents, like Tongyi DeepResearch-30B.

Paper Structure

This paper contains 44 sections, 4 equations, 5 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Deep Research agent systems.
  4. Data synthesis for deep research agents.
  5. Trajectory construction.
  6. Test-time scaling.
  7. Verified Data Synthesis
  8. Graph-Based Synthesis with Adversarial Verification
  9. Answer entity sampling.
  10. Structured attribute profiling.
  11. Reverse path search.
  12. Adversarial answer uniqueness verification.
  13. Agent-Based Web Exploration Synthesis
  14. Evidence-first question construction.
  15. Multi-stage quality verification.
  16. ...and 29 more sections

Figures (5)

  • Figure 1: Benchmark performance of our proposed Marco DeepResearch 8B-scale agent.
  • Figure 2: Overall framework of our verification-center design for Marco DeepResearch agents.
  • Figure 3: Distribution comparison between the open-source multi-hop QA dataset (2Wiki, BeerQA, etc.) and our synthesized data. Left: token count per sample. Right: tool-call rounds per sample.
  • Figure 4: Distribution comparison between three deep search data (ASearcher, DeepDive, and REDSearcher) and our synthesized data. Left: token count per sample. Right: tool-call rounds per sample. Our data is consistently shifted toward longer outputs and more tool interactions.
  • Figure 5: Relative to open-source data, our synthesized datasets indicate higher intrinsic difficulty.