Model Unlearning Objectives Vary for Distinct Language Functions

In three linesarXiv paper on selective unlearning in LLMs. Authors propose two distinct methods: a cosine-based RMU variant for dangerous-knowledge unlearning, and a multi-layer objective for toxicity reduction. Tested on 4 open-source 7-8B models, approaches show unlearning requires function-specific objectives, analogous to LLM post-training.

Summary generated by Claude — human-verified