Semantics-Based Analysis and Repair of Neural Language Models

This thesis treats foundation models not as mysterious black boxes, but as maintainable artifacts that engineers can inspect, intervene in, and improve over time. It develops three connected ideas: LM Analysis asks what a language model is doing and why; LM Repair corrects the model or its outputs when behavior goes wrong; and LM Meta-repair makes repair methods themselves safer and better. Across these steps, semantics is the operational medium, that is, human-understandable meaning becomes the practical handle for engaging with language models and shaping their mechanisms. The goal is to make foundation models more reliable, understandable, and useful.