Liu, C. (CSE) – Enabling LLM Unlearning at Inference Time by Decomposing Detection and Intervention

Machine unlearning addresses the “right to be forgotten” under GDPR and enables privacy, copyright, and safety compliance in large language models. Training-based unlearning can remove targeted behavior on benchmarks, but it scales poorly, can degrade utility, and can fail under adversarial prompting that recovers supposedly forgotten content. This prospectus proposes inference-time behavioral unlearning: rather than modifying weights to “erase” knowledge, we detect when a query targets forgotten content and intervene in generation so the system behaves like a model never trained on that content. We formalize this approach as Detect-Intervene Decomposition and instantiate it with three complementary methods operating at the embedding, token, and reasoning levels under different access capabilities. Comprehensive experiments across entity unlearning, hazardous knowledge removal, and copyright protection demonstrate that our methods match or exceed training-based approaches while being orders of magnitude faster and preserving model utility. As LLMs increasingly operate as services with restricted weight access, inference-time unlearning provides the only practical path for responsible AI deployment that respects privacy, safety, and legal requirements.

Event Host: Chris Liu, Ph.D. Student, Computer Science and Engineering

Advisor: Yang Liu

Zoom – https://ucsc.zoom.us/j/94799852992?pwd=EBFQe4U2lRNro1oJ8F36bgORhT2xSv.1

Passcode –  242384