Large language models (LLMs) have achieved outstanding performance across a wide range of natural language understanding and generation tasks. However, their internal mechanisms remain largely opaque, raising concerns about their trustworthiness. This...
Large language models (LLMs) have achieved outstanding performance across a wide range of natural language understanding and generation tasks. However, their internal mechanisms remain largely opaque, raising concerns about their trustworthiness. This dissertation presents a unified analytical framework—UniK (Unified Framework for Knowledge Localization)—that quantifies the contributions of each neuron to specific knowledge within LLMs. Building upon this foundation, the dissertation explores four major applications. First, Decomposition of Experts (DoE) introduces an unplug-and-play paradigm that dynamically deactivates task-irrelevant neurons, achieving inference speed-ups with minimal performance loss. Second, CRISPR mitigates biases in instruction-following models by pruning bias-related neurons, demonstrating that only a small subset of neurons encodes significant social or cognitive biases. Third, KLUE addresses privacy concerns by faithfully erasing sensitive knowledge through privacy-related neuron updates, supported by the FaithUn benchmark that considers knowledge interconnectedness. Finally, using the knowledge localization framework, we show that existing unlearning methods often hide rather than genuinely erase knowledge. To address this issue, we propose a regularization term that suppresses such hiding signals and promotes authentic forgetting. Collectively, these studies establish a clearer understanding of how task-specific, bias, and privacy knowledge is distributed within LLMs. These findings demonstrate that knowledge localization not only enhances interpretability but also enables proactive control over model behavior.