WBFT-MLLMN: A Weighted BFT Consensus Driven Multiple Large Language Models Network

Different Large Language Models (LLMs) give inconsistent and even different answers to the same questions. Therefore, we often query multiple LLMs for a more comprehensive and precise answer. To facilitate and speed up this process, a network of multiple LLMs, called multi-LLM Network (MLLMN), is born. It interacts with answers through pre-configured communication protocols. Unfortunately, different organizations' LLMs have unique perspectives on the same issues. Meanwhile, the devices loaded with these LLMs, such as mobile phones, personal computers, etc., may have malicious manipulation behavior. As a result, the accuracy and reliability of answers transmitted in MLLMN cannot be reliably guaranteed. To address the above problems, we have designed a Weighted Byzantine Fault Tolerance (WBFT) consensus that can accurately decide reliable answers in MLLMN without relying on third parties, even if there are malicious manipulators. The LLM weight is assigned based on the quality of the answers and the trust of LLMs, which will incentivize the LLM to respond reliably and avoid dishonest answering behavior as much as possible. The simulation and investigation results not only prove that WBFT consensus is superior to traditional consensus in efficiency and consensus success rate, in particular in the wireless environment, but also prove that MLLMN with the blessing of WBFT can obtain more accurate and reliable answers than a single LLM and MLLMN without consensus participation.