Composable Interventions for Language Models

Explore a comprehensive lecture on composable interventions for language models presented by Arinbjörn Kolbeinsson at the USC Information Sciences Institute. Delve into the world of test-time interventions that enhance factual accuracy, mitigate harmful outputs, and improve model efficiency without costly retraining. Discover a new framework for studying the effects of using multiple interventions on the same language models, featuring innovative metrics and a unified codebase. Examine extensive experiments composing popular methods from Knowledge Editing, Model Compression, and Machine Unlearning categories. Uncover meaningful interactions between interventions, including how compression affects editing and unlearning, the importance of application order, and the inadequacy of general-purpose metrics for assessing composability. Gain insights into clear gaps in composability and the need for new multi-objective interventions. Access the public codebase to further explore the concepts presented. Learn from Arinbjörn Kolbeinsson's expertise in responsible and accurate models for health and biomedicine, as well as his background in machine learning and biostatistics.