How can we really trust or certify AI if it does not give the same response when asked the same question twice? This talk will focus on consistency as a fundamental pillar of trust and interpretability in AI, both from theoretical and empirical perspectives including new measures of stability with theoretical guarantees, trustworthy data selection using similar stability measures, and new reward design for improving consistency in RAG systems. The next part of the talk will focus on few-shot distillation and miniature replicas for auditing, investigating how explainability techniques can help with faithful reconstruction of model behavior by precisely extracting the most relevant information.
[1] Quantifying Prediction Consistency Under Fine-Tuning Multiplicity in Tabular LLMs
[2] Few-shot knowledge distillation of llms with counterfactual explanations
Sanghamitra Dutta is an assistant professor in ECE along with affiliations at the Center for Machine Learning (CML) at UMIACS, the Department of Computer Science, and the AIM. Her research interests broadly revolve around reliable, efficient, and trustworthy machine learning, bringing in mathematical techniques from optimization, statistics, information theory, and causality. She received the 2024 NSF CAREER Award and was recently recognized in the 2025 List of 100 Brilliant Women in AI Ethics.