Accelerating Autonomous System Evaluation via Reinforcement Learning Driven Large Language Model Agents for Real Time Performance Diagnostics and Strategy Refinement

Abstract

The rapid proliferation of autonomous systems across critical infrastructures—ranging from transportation networks to industrial manufacturing—has outpaced traditional verification and validation methodologies. Conventional testing frameworks often rely on static scenarios that fail to capture the edge cases inherent in dynamic, real-world environments. This paper proposes a novel architectural paradigm for accelerating autonomous system evaluation by integrating Reinforcement Learning (RL) with Large Language Model (LLM) agents. This interdisciplinary approach leverages the high-level reasoning capabilities of LLMs to interpret complex system logs and the optimization prowess of RL to iteratively refine testing strategies in real time. By deploying these agents within a socio-technical framework, we provide a mechanism for continuous performance diagnostics and adaptive strategy refinement. The research emphasizes the structural trade-offs between computational latency and diagnostic depth, the governance of autonomous evaluators, and the long-term sustainability of AI-driven testing infrastructures. Our findings suggest that RL-driven LLM agents can significantly reduce the temporal requirements for identifying critical failure modes while enhancing the robustness and fairness of the autonomous systems under review. Furthermore, we discuss the policy implications of delegating safety-critical evaluation tasks to generative agents and propose a roadmap for integrating these systems into existing regulatory and engineering workflows.