PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2603.28376v1cs.CLcs.AI

TL;DR

Deep research agents lack explicit verification across QA synthesis, trajectory construction, and inference, risking error propagation during long-horizon tasks. Marco DeepResearch introduces verification-centric mechanisms across these stages and, as an 8B-scale agent, outperforms most 8B-scale agents while matching or surpassing several 30B-scale agents under bounded tool use.

  • Problem

    Current deep research agents lack explicit verification in QA data synthesis, trajectory construction, and inference, limiting reliability across long-horizon tasks.

  • Method

    Marco DeepResearch combines verified QA synthesis, verification-driven trajectory construction, and verifier-guided test-time scaling.

  • Results

    Marco DeepResearch significantly outperforms 8B-scale agents on most challenging benchmarks and, with up to 600 tool calls, surpasses or approaches several 30B-scale agents.

  • Takeaways & Limitations

    Verification-centric design enables the optimized 8B model to close part of the performance gap with larger deep research agents on complex web navigation and information-seeking tasks.

  • Takeaways & Limitations

    Agent-based synthesis can make data construction prohibitively expensive because it requires multi-turn web exploration.

Abstract

from arXiv · show

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.

1. Introduction

Marco DeepResearch addresses missing verification in QA synthesis, trajectory construction, and inference by introducing verification mechanisms across all three stages. The resulting 8B-scale agent outperforms most 8B-scale agents and approaches or surpasses several 30B-scale agents under bounded tool use.

  • Explicit verification is missing across QA data synthesis, trajectory construction, and inference, allowing errors to propagate and degrade agent performance.
  • Marco DeepResearch introduces verified QA synthesis, verification-driven trajectory construction, and verifier-guided test-time scaling.The agent itself serves as a verifier during inference under a controlled compute budget.
  • Marco DeepResearch is trained from Qwen3-8B and evaluated on six deep search benchmarks, including BrowseComp, BrowseComp-ZH, and GAIA.
  • Under a budget of up to 600 tool calls, the agent surpasses MiroThinker-v1.0-8B on BrowseComp-ZH and matches or exceeds several 30B-scale agents.

2. Related Work

Related work develops deep research agents, synthetic data, trajectory construction, and test-time scaling, but explicit verification remains underdeveloped across long-horizon information seeking. Existing synthesis paradigms especially struggle to produce difficult QA pairs with unique and correct answers.

  • Deep Research agent systems: Deep research agents build on agentic information seeking, autonomously planning, navigating multi-turn web interactions, and extracting reasoning-driven evidence.
  • Data synthesis for deep research agents: Graph-based synthesis traverses knowledge graphs, whereas agent-based synthesis explores real web environments to create multi-hop QA data.
  • Data synthesis for deep research agents: Both synthesis paradigms face the challenge of automatically generating difficult QA pairs with unique and correct answers.
  • Trajectory construction: ReAct-based and more advanced trajectory frameworks generally lack explicit verification during interactions, leaving intermediate results vulnerable to noise and misleading evidence.
  • Test-time scaling: Current test-time scaling emphasizes multi-agent coordination and context summarization, while the role of systematic explicit verification remains insufficiently developed.

3. Verified Data Synthesis

Verified Data Synthesis combines graph-based and agent-based pipelines with explicit checks for groundedness, uniqueness, correctness, and difficulty. Its adversarial and multi-stage verification procedures produce diverse, challenging QA data with fewer than 10% clear mismatches in a manual review of 100 samples.

  • Existing entity obfuscation can make multi-hop questions ambiguous, producing multiple valid answers and unstable or biased training data.
  • Graph-Based Synthesis with Adversarial Verification: Graph-based synthesis uses answer-first reverse construction and adversarial verification to jointly control search depth, uniqueness, correctness, and entity leakage.
  • Graph-Based Synthesis with Adversarial Verification: The graph pipeline samples constrained answer entities, profiles attributes, searches backward for 4 to 8 intermediate evidence nodes, and forms multi-hop reasoning chains.
  • Graph-Based Synthesis with Adversarial Verification: A Generator, Attacker, and Analyzer iteratively add constraints and search for counterexamples to verify answer uniqueness for at most 10 rounds.
  • Agent-Based Web Exploration Synthesis: Agent-based synthesis explores real-world web environments, constructs questions from verified evidence, and uses diverse reasoning topologies to control difficulty.
  • Agent-Based Web Exploration Synthesis: Multi-stage verification checks factual consistency, evidence grounding, retrieval necessity, reasoning depth, and alternative valid answers, with diagnosis-guided revision when samples fail.
  • Fewer than 10% of 100 manually reviewed samples had clear question-answer mismatches, while the remaining samples were valid but challenging.

4. Verification-Driven Trajectory Construction

Verification-Driven Trajectory Construction inserts explicit checks for intermediate sub-task outputs and final answers into training trajectories. It uses independent verification and feedback-based rerollout to correct failed trajectories before retaining them.

  • Standard single-agent ReAct trajectories often omit explicit checks, allowing early errors to propagate and accumulate into degraded final performance.
  • Multi-agent with Verification: A three-role framework assigns problem decomposition and aggregation to a main agent, sub-task solving to a search sub-agent, and independent web validation to a verifier.
  • Multi-agent with Verification: When verification fails, the corresponding step is revised and re-executed, recording verification-driven correction behavior in the trajectory.
  • Verification-Reflection Re-rollout on Failed Trajectories: Failed trajectories are diagnosed by a verifier, rerun using actionable feedback, and retained when the final answer is recovered correctly.

5. Verifier-Guided Test-Time Scaling

Marco DeepResearch scales inference by verifying candidate answers and resetting degraded trajectories, rather than relying only on more interaction rounds. These complementary mechanisms improve trajectory quality and answer quality without changing model parameters.

  • Discard All: Discard All removes accumulated tool history and intermediate reasoning after degeneration signals, then restarts from the original query and system prompt.This reset enables new search paths and reduces error propagation within a trajectory.
  • Verifier-Guided Test-time Scaling: Verifier-Guided Test-time Scaling checks candidate answers with rule-based checks and Marco DeepResearch as an agent judge.Candidates are independently verified while exploration continues, then jointly verified when the process converges or reaches the maximum step budget.
  • Combined Design: The two components are complementary: Discard All improves trajectory quality, whereas verifier-guided scaling improves answer quality.Together they enable more effective test-time scaling and stronger inference-time gains on hard questions.

6. Training Pipeline

The training pipeline combines supervised fine-tuning with reinforcement learning, using masked assistant-token loss, group-relative policy optimization, and staged outcome judging.

  • Training Pipeline: The training pipeline consists of Supervised Fine-Tuning and Reinforcement Learning.
  • Supervised Fine-Tuning: Supervised fine-tuning uses token-level cross-entropy while masking instruction and tool-response content from optimization.Only assistant response tokens contribute to the supervised loss.
  • Reinforcement Learning: Group Relative Policy Optimization updates the policy using advantages computed from relative rewards within each group.The importance sampling ratio compares the current policy with the old policy.
  • Reinforcement Learning: Relative advantage is computed by normalizing rewards within each group.
  • Reward Design: A two-stage LLM-as-Judge pipeline uses Qwen-Turbo-Latest for all samples and escalates uncertain or low-confidence cases to GPT-4.1.The outcome-based reward balances reward quality and computational cost.

7. Experimental Setup

Marco DeepResearch is evaluated across six deep-search benchmarks against foundation models with tools and trained agents at or above 30B scale. The setup also specifies verified and open-source training data alongside an efficiency-oriented implementation.

  • Benchmarks: The evaluation covers six benchmarks: BrowseComp, BrowseComp-ZH, GAIA, xBench-DeepSearch, WebWalkerQA, and DeepSearchQA.They assess information seeking, agentic information seeking in Chinese, real-world multi-step questions, diverse-domain search, web navigation, and exhaustive answer-set generation.
  • Baselines: Baselines include tool-using foundation models and trained agents at or above 30B scale.The listed comparison groups include models such as OpenAI-o3, GPT-5 High, Gemini-3-Pro, Tongyi DeepResearch, WebSailor-v2, and MiroThinker variants.
  • Training Data: Training data combines open-source QA and search datasets with synthetic real-world e-commerce and verified data.The open-source sources include 2Wiki-MultihopQA, BeerQA, ASearcher, DeepDive, QA-Expert-Multi-Hop-QA, and REDSearcher.
  • Implementation Details: Implementation uses Qwen3-8B with a 128K context window extended by YaRN and training on 64 A100 GPUs.The system also uses caching, retries, asynchronous tool calls, asynchronous reward computation, and synchronized deployment for efficiency and stability.

8. Experimental Results

Marco-DeepResearch-8B outperforms other 8B-scale agents on most deep-search benchmarks and remains competitive with larger systems. Analyses attribute gains to difficult verified data, verification-driven trajectories, reinforcement learning, verifier-guided test-time scaling, and longer-context training.

  • Main Results: Marco-DeepResearch-8B achieves the highest scores among its size category on BrowseComp (31.4), BrowseComp-ZH (47.1), WebWalkerQA (69.6), and xBench-DeepSearch (82.0 on 2505; 42.0 on 2510).
  • Main Results: Marco-DeepResearch-8B remains competitive on the remaining benchmarks, missing the GAIA text-only top score by 0.5 points and approaching or surpassing several 30B-scale agents.
  • Data Statistics Analysis: Our synthesized samples contain longer token sequences and more tool-call rounds than existing multi-hop and deep-search datasets.The reported shift is intended to provide denser supervision for maintaining state, revising hypotheses, and aggregating evidence across extended tasks.
  • Data Statistics Analysis: 29.0% < 51.7%: under identical ReAct-style construction with the same frontier agent, generated data has a lower answerable rate than open-source data.This indicates a harder data distribution.
  • Effect of QA Data Verification: Adversarial uniqueness verification improves downstream performance across most benchmarks by filtering noisy and ambiguous samples before trajectory construction and training.The comparison uses graph-based synthesis pipelines with identical data scales, differing only in whether verification is applied.
  • Verification-Driven Trajectory Construction: +2.03 average improvement: augmenting single-agent ReAct trajectories with multi-agent verified trajectories consistently improves performance across all benchmarks.
  • Improvement of Reinforcement Learning: +2.6 points average gain: reinforcement learning improves performance across all five benchmarks, with gains ranging from +0.8 to +6.7 points.
  • Test-Time Scaling: +12.1 points average gain: verifier-guided test-time scaling improves performance across the reported BrowseComp and BrowseComp-ZH settings under expanded tool-call budgets.

9. Conclusion

The paper presents Marco DeepResearch as an 8B-scale agent built around verification-centric design across data synthesis, trajectory construction, and inference. Experiments and ablations report improvements over smaller agents and several larger systems on challenging benchmarks.

  • Conclusion: Marco DeepResearch addresses missing verification in QA synthesis, trajectory construction, and inference through a unified verification-centric design.The three components are verified QA synthesis, verification-driven trajectory construction, and verifier-guided test-time scaling.
  • Conclusion: The agent significantly outperforms 8B-scale open-source deep research agents on most BrowseComp and BrowseComp-ZH benchmarks.
  • Conclusion: Marco DeepResearch surpasses several 30B-scale deep research agents on BrowseComp-ZH.
  • Conclusion: Detailed analysis and ablation studies support positive contributions from the proposed verification-centric designs.

Contributions and Acknowledgements

The acknowledgements recognize the collaborative effort behind Marco DeepResearch and identify the project lead and contributor groups. The listed contributors include core contributors and additional contributors.

  • Acknowledgements: The paper describes Marco DeepResearch as a highly collaborative effort involving all team members.
  • Acknowledgements: Longyue Wang is identified as the project lead.
  • Acknowledgements: Bin Zhu, Qianghuai Jia, Tian Lan, Junyang Ren, Feng Gu, Feihu Jiang, Longyue Wang, Zhao Xu, and Weihua Luo are listed as core contributors.
  • Acknowledgements: Yu Zhao, E. Zhao, Jingzhen Ding, Yuxuan Han, ChenLin Yao, Jianshan Zhao, Wanying Chen, Jiahong Wang, Jiahe Sun, Wanghui huang, Yongchao Ding, Junyuan Luo, Junke Tang, Zhixing Du, Zhiqiang Yang, Haijun Li, and Huping Ding are listed as contributors.
Loading 2603.28376v1…