This paper proposes collaboration among multiple LLM agents to overcome the limitations of a single LLM agent (hallucination and single point of failure). In particular, as LLM agents are increasingly deployed on open blockchain platforms, a multi-agent system that can tolerate malicious (Byzantine) agents is essential. Existing Byzantine-resistant multi-agent systems rely on leader-driven coordination, which is vulnerable to targeted attacks on the leader, and the leader's low-quality proposal can be accepted as the final answer even when there are high-quality alternatives. To address this, this paper proposes DecentLLMs, a decentralized consensus approach in which worker agents generate answers simultaneously and evaluation agents independently evaluate and rank the answers to select the best answer. DecentLLMs enables faster consensus despite the presence of Byzantine agents and consistently selects higher-quality answers through Byzantine-resistant aggregation techniques. Experimental results show that DecentLLMs effectively tolerates Byzantine agents and significantly improves the quality of selected answers.
