Title: Ev-Trust: Establishing Evolutionarily Stable Trust in Decentralized LLM-Driven Multi-Agent Service Markets

Abstract:

Traditional trust frameworks are increasingly vulnerable in decentralized service economies powered by multi-agent Large Language Models (LLMs). These systems suffer from three critical weaknesses: the lowered barrier to committing fraud, the challenge of accurately assessing service quality, and the inherent instability of service outputs. Together, these factors can precipitate a systemic collapse of trust across the population and encourage the spread of short-term, self-serving strategies.

To counter these challenges, we introduce Ev-Trust, a novel mechanism designed to be evolutionarily stable. Our approach incorporates three specific innovations: first, a cross-validation gate that utilizes the requestor’s semantic understanding to verify the validity of responses; second, a variance-standardized drift measure that distinguishes genuine behavioral anomalies from endogenous stochastic noise; and third, the integration of trust metrics into the expected revenue function, thereby transforming trustworthiness into a tangible evolutionary survival advantage.

Using replicator dynamics grounded in a noisy best-response micro-foundation, we demonstrate the asymptotic stability of cooperative evolutionarily stable strategies and establish explicit threshold conditions necessary to sustain cooperative equilibria. We validated Ev-Trust through simulations comprising 100 rounds, involving a minimum of 100 heterogeneous LLM-driven agents representing seven distinct behavioral types. The testing environment utilized TruthfulQA and TriviaQA, two standard benchmarks for factual question-answering.

When compared against baseline models relying on transitive trust aggregation, reinforcement-learning-based reputation systems, and pure evolutionary imitation, Ev-Trust demonstrated significant improvements. Specifically, it reduced the participation of malicious agents by roughly 60% and lowered the rate of fraudulent services by approximately 50%. Furthermore, the mechanism preserved stable trust differentiation even when subjected to a 30% adversarial mutation rate. These findings indicate that combining semantic-based trust evaluation with evolutionary incentives offers a robust, principled framework for ensuring cooperation within decentralized LLM-based multi-agent ecosystems.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC