Date of Award

2026

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Science

Committee Chair

Tathagata Mukherjee

Committee Member

Letha Hughes Etzkorn

Committee Member

Jacob Hauenstein

Research Advisor

Tathagata Mukherjee

Subject(s)

Artificial intelligence--Data processing, Natural language processing (Computer science), Computer programs--Correctness, Counterfactuals (Logic)

Abstract

Correctness and faithfulness are standard metrics used for evaluating Retrieval-Augmented Generation (RAG) systems, but they only describe the outcomes and do not surface the underlying behaviors behind them. The same scores can hide a number of underlying behaviors. They can hide whether a system relied on retrieval, answered from parametric knowledge, followed a near-miss context, missed the gold source or failed to use the retrieved evidence. The same scores can hide all of these. This thesis introduces a counterfactual behavioral evaluation framework that tests each query under five controlled context conditions and uses a gold-source coverage gate to identify whether the annotated source passage was retrieved. These observations together are used to assign each query to one of the ten behavioral labels defined in this thesis, which indicate to a RAG practitioner where engineering attention should be directed. The framework has been evaluated across three diverse knowledge domains and two model capability classes. The framework reveals distinctions that are hidden by correctness and faithfulness alone and it treats these metrics as starting points for RAG diagnosis and not as endpoints.

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.