Quality Assurance in Large Language Models

Large Language Models (LLMs) have demonstrated impressive capabilities across a wide range of tasks in natural language processing (NLP) and software engineering (SE). However, their increasing adoption has raised critical concerns about trustworthiness, including robustness under distribution shifts, vulnerability to hallucinations, and misalignment between confidence and accuracy. These issues are exacerbated by the complex and probabilistic nature of LLMs, which makes it difficult for developers to observe, interpret, and address their failures. To bridge this gap, this thesis presents a series of tools and frameworks that enable a more transparent and actionable understanding of LLM trustworthiness. Together, these works follow a methodological arc that moves from interactive exploration to internal structure modeling and finally to external behavioral correction. The first contribution, DeepLens, introduces a human-centered system that supports interactive diagnosis of out-of-distribution (OOD) data in large text corpora. By integrating clustering-based analysis and neuron-level highlighting, {\deeplens} empowers developers to detect, categorize, and interpret OOD issues that threaten model reliability. Building on this foundation, the second contribution, LUNA, proposes a white-box framework that models LLM behavior using abstract probabilistic structures. By mapping hidden states to interpretable semantic concepts, LUNA enables structured evaluation of trustworthiness dimensions such as hallucination and adversarial robustness, providing insights grounded in both semantic and model-level perspectives. The final contribution addresses the limitations of white-box methods in black-box scenarios. It introduces a confidence calibration framework that quantifies and corrects model confidence based on semantic-preserving behavioral variations. By leveraging agreement ratios, entropy, and stability features from multiple LLM outputs, this method learns a regression-based correction factor to improve alignment between model confidence and factual correctness, without requiring model internals. Together, these three works provide a practical toolchain to enhance the transparency and trustworthiness of LLMs, spanning interactive user interfaces, interpretable model-based analysis, and scalable confidence calibration.