Please note: This PhD seminar will take place in DC 3301.
Joel Rorseth, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Lukasz Golab
Retrieval-augmented generation (RAG) has become a key component of modern large language model (LLM) systems due to its ability to integrate external knowledge sources. However, this integration makes it difficult to determine which retrieved sources actually influence a model’s output.
To address this challenge and recover provenance in RAG systems, we present a framework for explaining how individual sources shape LLM responses. First, we introduce counterfactual explanations that identify how different source subsets lead to different outputs. Building on this foundation, we propose if-then rules that capture consistent relationships between source subsets and specific output conditions, thereby summarizing broader provenance patterns. To compute these explanations efficiently, we develop lattice-based search algorithms that incorporate novel pruning strategies. Finally, we review experimental results demonstrating both the efficiency gains achieved through these optimizations and the practical utility of the resulting provenance explanations.