Graph-Based Evidence Selection for Improving Retrieval-Augmented Generation in Large Language Models

Large language models increasingly rely on retrieval-augmented generation (RAG) to supplement their internal knowledge with external documents, reducing hallucinations and allowing responses to incorporate current information. Despite these advantages, retrieval systems generally assume that the highest-ranked documents are the most useful inputs for generation. In practice, retrieved passages frequently contain redundant evidence, conflicting claims, or information that is only loosely related to the user's question. As more documents are added to the context window, model performance often plateaus or even declines because attention becomes divided across increasingly noisy inputs. Current retrieval pipelines therefore optimize for document relevance rather than for the quality of the complete evidence set presented to the language model.

This project investigates whether retrieval can be improved by optimizing the diversity and complementary information content of retrieved evidence rather than maximizing similarity scores alone. Instead of selecting the top k passages directly from a dense retriever, the system constructs a graph in which nodes represent candidate documents and weighted edges encode semantic overlap between them. A graph optimization procedure then identifies subsets of documents that collectively maximize topic coverage while minimizing redundancy. These curated evidence sets are passed to a downstream language model for question answering, with the central hypothesis that more diverse supporting information will improve factual accuracy, calibration, and reasoning performance without requiring additional retrieved context.

The group will evaluate the framework across established open-domain question answering benchmarks using publicly available Wikipedia-based corpora and compare graph-based evidence selection against conventional dense retrieval, maximal marginal relevance, and reranking approaches under identical retrieval budgets. Evaluation will measure answer correctness, citation precision, hallucination frequency, and robustness under adversarial retrieval scenarios containing partially conflicting evidence. Beyond overall performance, the study will examine how graph topology changes across different question types and whether certain structural properties of the selected evidence subgraphs correlate with improved reasoning.