Uncertainty-Aware Reinforcement Learning for Robust Decision Making in LLM-Agent Collaboration

Abstract

Decision-making processes involving LLM agents are often susceptible to uncertainty arising from ambiguous inputs and stochastic generation behavior. To address this issue, this study proposes an uncertainty-aware reinforcement learning framework that incorporates Bayesian reward estimation and entropy-regularized policy updates. The approach is validated on a benchmark consisting of 9,600 tasks with varying levels of input ambiguity, including incomplete instructions and conflicting objectives. Results indicate that the proposed method reduces error propagation by 28.7% and improves decision robustness, with task success rates increasing from 68.9% to 80.2%. Additionally, calibration metrics such as expected calibration error (ECE) decrease by 19.5%, demonstrating improved reliability in agent outputs. The framework provides a systematic solution for enhancing robustness in LLM-based collaborative systems under uncertainty.