Mutual Influence AI: Trust-Based Cooperation Mechanisms for LLM Multi-Agent Systems

Abstract

This paper introduces Mutual Influence AI, a novel concept for adaptive cooperation in multi-agent systems. Unlike classical independent reasoning or cen- tralized orchestration, our approach introduces an ex- plicit mutual influence factor μ that captures trust- adjusted peer feedback and directly modulates large lan- guage model (LLM) generation. We present (i) a math- ematical formalization of mutual influence, (ii) a pro- totype implementation integrated with Microsoft Auto- Gen for LLM-based agents, and (iii) qualitative evi- dence that the framework improves adaptability, trans- parency, and coordination in multi-agent dialogues. Results show that Mutual Influence AI stabilizes group interactions efficiently while providing interpretable control over how agents influence each other. This positions Mutual Influence AI as a new paradigm for LLM-driven multi-agent systems with potential appli- cations ranging from collaborative problem solving to cybersecurity. Quantitatively, across 167 simulation runs, cross–role agreement increased from 0.19 (base- line) to 0.50 under influence (approx. +160%), with median revision depth (approx. 1.0). Under adversarial feedback, agreement still improved (0.18 to 0.47).