Reproduce “MIRROR: Multi-agent Intra- and Inter-Reflection for Optimized Reasoning in Tool Learning” on OpenResearch

About this paper

Zikang Guo, Benfeng Xu, Xiaorui Wang, Zhendong Mao

The MIRROR framework, developed by researchers at the University of Science and Technology of China and Metastone Technology, proposes a multi-agent system integrating novel intra-reflection (pre-execution self-assessment) with inter-reflection (post-execution learning). This dual-reflection approach enables proactive error prevention and continuous learning, leading to superior performance on challenging tool-use benchmarks, including an 83.7% Pass Rate on StableToolBench with GPT-3.5 Turbo and a 100% Delivery Rate on TravelPlanner with GPT-4o.

Key insights

• Proactive intra-reflection by individual agents is crucial for preventing errors before execution, significantly improving reliability and minimizing error propagation in multi-agent systems.
• The synergistic integration of intra-reflection (pre-action self-assessment) and inter-reflection (post-execution learning) creates a robust and efficient error prevention and correction system for complex tool-use tasks.
• Smaller, optimized LLM variants can achieve high performance in complex tool-use tasks when integrated into sophisticated reflection frameworks like MIRROR, suggesting that performance is not solely dictated by model size.

Summary by alphaXiv.