Improving LLM Reasoning and Retrieval for Structured and Complex Information Spaces

Abstract

Large Language Models (LLMs) excel at fluent language generation but face critical challenges in high-stakes domains that require reasoning over long contexts, structured information use, grounded retrieval, and human-verifiable outputs. This dissertation explores how to improve LLM performance on complex, context-rich tasks through four contributions. First, we introduce memory-augmented architectures for multi-document reasoning, highlighting gaps between summarization and true inference. Second, we benchmark relational reasoning by reconstructing latent graphs from long texts, revealing a limitation we term "memory drift." Third, we show that incorporating structured metadata as a first-class signal in retrieval-augmented generation (RAG) systems improves retrieval consistency in large, repetitive corpora by better disambiguating context. Finally, we present a human-in-the-loop system for structured data analysis that enables transparent, code-centric interaction and supports iterative sensemaking over complex datasets. Together, these efforts advance LLM capabilities in analytical synthesis, structured retrieval, long-context evaluation, and explainability, offering practical tools for building more trustworthy and effective AI systems in real-world applications.