Weak-Link Optimization for Multi-Agent Reasoning and Collaboration

arXiv:2604.15972v1 Announce Type: new Abstract: LLM-driven multi-agent frameworks address complex reasoning tasks through multi-role collaboration. However, existing approaches often suffer from reasoning instability, where individual agent errors are amplified through collaboration, undermining overall performance. Current research mainly focuses on enhancing high-capability agents or suppressing unreliable outputs to improve framework effectiveness, while systematic identification and reinforc