Reproduce “AgentCDM: Enhancing Multi-Agent Collaborative Decision-Making via ACH-Inspired Structured Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xuyang Zhao, Shiwan Zhao, Hualong Yu, Liting Zhang, Qicheng Li
Researchers at Nankai University developed AgentCDM, a two-stage training framework for multi-agent LLMs that incorporates an Analysis of Competing Hypotheses (ACH)-inspired structured reasoning protocol to enhance collaborative decision-making. This system consistently achieved state-of-the-art performance, outperforming single-agent baselines by up to 29.3 percentage points on benchmarks like MMLU-PRO, and demonstrated strong generalization across diverse datasets.
Key insights
- Implementing an ACH-inspired structured reasoning protocol within LLM-based multi-agent systems effectively mitigates cognitive biases and transforms decision-making from passive answer selection to active hypothesis evaluation and construction.
- A two-stage training approach, starting with explicit scaffolding and gradually transitioning to autonomous exploration, is crucial for LLMs to internalize structured reasoning capabilities and achieve robust generalization.
- Training on more challenging datasets can cultivate deeper, more transferable reasoning abilities that generalize exceptionally well across diverse tasks and domains, highlighting the importance of curriculum design for robust AI.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.