Event Details

Speaker: Arinbjörn Kolbeinsson, University of Virginia

Virtual zoom link: https://usc.zoom.us/j/7042850182?pwd=OTQ3aW9LUjErTC9iWGRFQUg0LzlOdz09&omn=96405030645
Meeting ID: 704 285 0182
Meeting password: 832239

Abstract: Test-time interventions for language models can enhance factual accuracy, mitigate harmful outputs, and improve model efficiency without costly retraining. But despite a flood of new methods, different types of interventions are largely developing independently. In practice, multiple interventions must be applied sequentially to the same model, yet we lack standardized ways to study how interventions interact. We fill this gap by introducing composable interventions, a framework to study the effects of using multiple interventions on the same language models, featuring new metrics and a unified codebase. Using our framework, we conduct extensive experiments and compose popular methods from three emerging intervention categories — Knowledge Editing, Model Compression, and Machine Unlearning. Our results from 310 different compositions uncover meaningful interactions: compression hinders editing and unlearning, composing interventions hinges on their order of application, and popular general-purpose metrics are inadequate for assessing composability. Taken together, our findings showcase clear gaps in composability, suggesting a need for new multi-objective interventions. All of our code is public: https://github.com/hartvigsen-group/composable-interventions

Hosted by: Abel Salinas
POC: Maura Covaci

Speaker Bio

Arinbjörn Kolbeinsson is currently serving as a visiting scholar at the University of Virginia, focusing on responsible and accurate models for health and biomedicine. His recent research explores the editing and efficiency of language models, along with the development of composable intervention techniques. Previously, Arinbjörn was a machine learning scientist at Evidation Health Inc., where he developed innovative methods to predict health outcomes using high-frequency multi-modal data. His work was pivotal in advancing differential privacy, disease modeling, and reinforcement learning for health applications. Arinbjörn completed his Ph.D. in Biostatistics at Imperial College London in 2020, specializing in deep learning for health outcome prediction.