A Closed-Loop Framework for Trustworthiness Assessment in Large Language Models

Abstract

: Trustworthiness is a fundamental requirement for Large Language Models (LLMs), particularly in high-stakes domains such as healthcare, finance, and education, where unreliable model behavior may lead to incorrect or misleading responses. We distinguish safety , which concerns harmful or undesirable outputs, from security , which concerns vulnerabilities arising from adversarial manipulation of model inputs or behavior. This work focuses on robustness and prediction stability under adversarial prompting . Despite strong performance across a wide range of tasks, LLMs remain vulnerable to adversarial prompting , where small semantically equivalent perturbations in the input can produce unstable or incorrect predictions. Existing approaches address this problem through robustness training, alignment with human feedback, or calibration approaches; however, they often focus on individual aspects of model behavior and do not adequately examine prediction stability under perturbations. To address this limitation, we propose Closed-Loop Trustworthiness (CLoT), an inference-time framework for improving robustness and prediction stability under adversarial prompting . The framework consists of two stages: (i) Evaluation and Failure Detection (Stage I) , which analyzes model responses under controlled perturbations to identify inconsistent predictions associated with unreliable behavior; and (ii) Evaluation-Guided Intervention (Stage II) , which applies perturbation-consistent post-hoc correction without modifying model parameters. We empirically evaluate CLoT on six benchmarks (SQuAD 2.0, AdvBench, HH-RLHF, HotpotQA, FEVER, and TruthfulQA) and observe performance degradation of up to 28% under adversarial prompting . Experimental results show that CLoT improves robust accuracy by up to + 4.9% and failure detection performance by up to + 12.5%, while reducing calibration error (e.g., ECE to 0.050).