Probing Hidden Knowledge Holes in Unlearned LLMs

Ming Jin (Griffith University) · Myeongseob Ko (Virginia Polytechnic Institute and State University) · Hoang Anh Just (Virginia Tech) · Charles Fleming (Cisco) · Ruoxi Jia (Virginia Tech)
content removalevaluation frameworkhidden costsirrelevant responsesknowledge holesknowledge preservationmachine unlearningmodel evaluationnonsensical responsesperformance benchmarkspre-trainingstatic benchmarkstest case generationunintended lossesunlearning techniques

Machine unlearning has emerged as a prevalent technical solution for selectively removing unwanted knowledge absorbed during pre-training, without requiring full retraining. While recent unlearning techniques can effectively remove undesirable content without severely compromising performance on standard benchmarks, we find that they may inadvertently create ``knowledge holes''