Robust Multi-Agent LLMs under Byzantine Faults
Researchers have introduced Self-Anchored Consensus (SAC), a new protocol designed to enhance the reliability of decentralized Large Language Model (LLM) multi-agent systems. As LLM agents increasingly collaborate via peer-to-peer networks, they become vulnerable to Byzantine faults, where unreliable or adversarial agents manipulate neighbors into incorrect conclusions. Existing solutions, such as leader-based coordination, are susceptible to such manipulation. The proposed SAC protocol operates as a fully decentralized, iterative filter-and-refine mechanism. Agents exchange responses, locally evaluate and filter out unreliable messages, and refine their outputs based on robust communication graph conditions. Experimental results on mathematical and commonsense reasoning benchmarks demonstrate that SAC effectively suppresses Byzantine influence and improves performance across various network topologies. In contrast, prior methods significantly degrade under adversarial conditions. This study addresses critical security vulnerabilities in collaborative AI systems, offering a pathway for more resilient decentralized artificial intelligence applications without relying on central authorities or self-reported confidence metrics.
Wire timeline
Robust Multi-Agent LLMs under Byzantine Faults
Researchers have introduced Self-Anchored Consensus (SAC), a new protocol designed to enhance the reliability of decentralized Large Language Model (LLM) multi-agent systems. As LLM agents increasingly collaborate via peer-to-peer networks, they become vulnerable to Byzantine faults, where unreliable or adversarial agents manipulate neighbors into incorrect conclusions. Existing solutions, such as leader-based coordination, are susceptible to such manipulation. The proposed SAC protocol operates as a fully decentralized, iterative filter-and-refine mechanism. Agents exchange responses, locally evaluate and filter out unreliable messages, and refine their outputs based on robust communication graph conditions. Experimental results on mathematical and commonsense reasoning benchmarks demonstrate that SAC effectively suppresses Byzantine influence and improves performance across various network topologies. In contrast, prior methods significantly degrade under adversarial conditions. This study addresses critical security vulnerabilities in collaborative AI systems, offering a pathway for more resilient decentralized artificial intelligence applications without relying on central authorities or self-reported confidence metrics.
cs.AI updates on arXiv.org