Researchers have proposed a new framework called RUPA (Relational Uncertainty Propagation for Agents) to improve uncertainty quantification in large language model agents. This is important because existing methods often rely on local signals and fail to identify errors that accumulate over time. RUPA represents an execution history as a graph, propagates uncertainty through the graph, and combines it with other features to produce a confidence estimate for the full agent trajectory. The authors evaluate RUPA on various benchmarks and find that it outperforms existing methods in providing accurate uncertainty estimates, earlier failure detection, and improved agent execution. --- Why it matters: This matters because reliable uncertainty quantification is crucial for deploying AI agents in complex environments where errors can accumulate over time. By accurately estimating uncertainty, developers can improve the trustworthiness of their agents and prevent failures that might have serious consequences. Source: https://arxiv.org/abs/2608.16002