Dimensions overall score 7.0

GitHub Code Pulse

No public code linked for this paper yet.

• Marco DeepResearch outperforms 8B-scale deep research agents on most challenging benchmarks, including BrowseComp and BrowseComp-ZH.result 95%
• Under a maximum budget of 600 tool calls, Marco DeepResearch surpasses or approaches several 30B-scale agents, like Tongyi DeepResearch-30B.result 90%
• The verification-centric framework introduces verification mechanisms to graph-based and agent-based QA synthesis to control question difficulty and ensure answer uniqueness and correctness.method 90%
• A verification-driven trajectory synthesis method injects explicit verification patterns into training trajectories.method 85%
• Marco DeepResearch itself is used as a verifier at inference time to improve performance on challenging questions.method 90%
• Existing agent frameworks share a critical limitation: the absence of explicit verification during interactions.limitation 80%
• The synthesized QA data for Marco DeepResearch is shifted toward longer outputs and more tool interactions compared to other deep search datasets.technical 85%
• Marco DeepResearch achieves competitive performance on exploration-heavy tasks, scoring 31.4 on BrowseComp, 47.1 on BrowseComp-ZH, 69.6 on WebWalkerQA, and 82.0 on xBench-DeepSearch (2505 split).result 95%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card