CCoRe: Cooperative-Competitive Reasoning LLM-based Multi-Agent Framework

Abstract Large Language Model-based Multi-Agent systems (LLM-MAS) emerged as a promising approach for solving complex tasks and queries that go beyond Single-Agent systems’ abilities. Cooperative-Competitive Reasoning LLM-based Multi-Agent Framework (CCoRe) is an open-source framework that allows developers to build Question Answering applications using lightweight LLM-MAS approach. These agents can converse with each other internally through DeepMonologue to accomplish complex user tasks by following one top-voted TaskGraph. Existing LLM-MAS can already solve simple dialogue tasks. CCoRe WiserAgents are topic customizable and can operate in two agent modes, Single-Agent (SA) or Multi-Agent (MA) mode, to better handle hard and easy queries, employing combinations of Large Language Models (LLMs), Human-In-The-Loop (HITL) and tools such as, Web Search and Wikipedia Search. Empirical studies demonstrate the framework’s effectiveness in many domains including commonsense reasoning mathematics and coding. The results show that CCoRe outperforms mid-to-heavyweight LLMs in GQC scores on the CRITICBENCH benchmark by 8.27\%, 13.74\% and 9.34\% in Generation (G), Critique (Q) and Correction (C) scores respectively. Additionally, reduced hallucination and lower resource consumptions are observed.