DuMate-DeepResearch: An Auditable Multi-Agent System with Recursive Search and Rubric-Grounded Reasoning

Researchers have introduced DuMate-DeepResearch, a multi-agent AI system designed to handle complex research tasks with improved auditability and reasoning. The framework achieves state-of-the-art results on deep research benchmarks by combining dynamic planning, recursive task delegation, and rubric-based quality optimization.

DuMate-DeepResearch addresses fundamental challenges in autonomous research systems by reimagining how AI agents approach open-ended, complex inquiries. Traditional deep research systems struggle with maintaining coherence across long planning horizons, managing computational complexity, and preventing hallucination during synthesis—problems that compound when tasks require iterative evidence gathering and verification. This work solves these issues through architectural decomposition, separating the planning and scheduling layer from execution tools, which enables transparency at every decision point.

The system's three core innovations represent meaningful advances in agentic AI. The graph-based dynamic planning strategy allows continuous roadmap refinement through reflection and parallel exploration, addressing the brittleness of fixed task decomposition. The recursive two-level execution delegates search subtasks to specialized agents, isolating noise and improving stability—a pattern increasingly valuable as systems scale complexity. Rubric-based test-time optimization is particularly significant: generating task-specific quality criteria as live reasoning scaffolds directly tackles hallucination risk by grounding synthesis in evidence.

For the broader AI ecosystem, DuMate-DeepResearch demonstrates that multi-agent architectures with explicit auditability can outperform monolithic approaches. Achieving 58% and 61.95% scores on respective benchmarks while ranking first in information recall suggests practical applicability beyond research tasks. This work influences how enterprises deploy AI for knowledge work, as auditability and reasoning transparency become critical for compliance and trust.

The framework's foundation on Qianfan Agent Foundry indicates growing infrastructure maturity for complex agentic systems. Future developments will likely focus on scaling these mechanisms to real-world research at enterprise scale, with particular attention to how rubric-based optimization reduces domain-specific tuning overhead.

• →DuMate-DeepResearch achieves state-of-the-art benchmarks (58-61.95%) by decoupling planning from execution and maintaining full auditability
• →Recursive two-level execution with specialized Search Agents improves stability for long-horizon complex tasks
• →Rubric-based test-time optimization dynamically generates quality criteria to ground synthesis and reduce hallucination
• →Graph-based dynamic planning enables continuous refinement through reflection, backtracking, and parallel branching
• →Multi-agent architecture with explicit traceability addresses enterprise requirements for transparency and compliance in autonomous research