Multi-agent systems (MAS) powered by LLMs frequently fail, and current reflection methods force all agents to reflect on failures, which contaminates the memory of agents that behaved correctly. This paper introduces DoCtOR, a framework that first automatically attributes the failure to the decisive error agent (the one that led the task astray) using a process reward-based method (ProFA). It then generates a corrected decisive error step via and has only that agent reflect. Experiments on HotPotQA, ChartQAPro, and Mind2Web show DoCtOR improves success rates by 22%, 26%, and 27% respectively, outperforming existing methods like Reflexion and COPPER. This approach is interesting because it addresses a fundamental oversight in MAS self-reflection, potentially making more reliable.
Targeted reflection improves multi-agent systems by locating decisive errors
DoCtOR attributes a failed trajectory to a specific agent and step, generates a counterfactual correction, and asks only that agent to reflect.
Chinese Tech
Xiaoqing Wang · Keman Huang · Bin Liang · Hongyu Li · Xiaoyong Du · Wuqiong Pan
Renmin University of China · Ant Group
Research Digest··2 min read
Thread:Multi-Agent Coordination
In one line
Multi-agent systems improve 22-27% by identifying and reflecting only the agent that caused the failure.
Before you read this
Three ideas this paper builds onfailure attributionIdentifying which component or step in a system caused an observed failure, often using structured trace representations.115 papers →multi-agent collaborationMultiple autonomous agents coordinate and combine their specialized capabilities to solve complex tasks beyond individual ability.16 papers →agent memoryA repository where an agent stores past interactions and outcomes to inform future decisions.200 papers →
Read it as
Wang et al. developed DoCtOR, a reflection framework that diagnoses which agent and reasoning step decisively caused a multi-agent task failure before generating corrective feedback.
The authors built DoCtOR, short for Diagnose-then-Correct PPO-enhanced Reflection.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.