ICCU: In-Context Continual Unlearning Framework for Language Models

A new framework called ICCU (In-Context Continual Unlearning) has been introduced by researchers to handle sequential unlearning requests in language models without altering model parameters. Unlike traditional fine-tuning approaches, ICCU generates understandable refusal rules from unlearning datasets, which are then utilized as a filter or through system prompts during inference. This framework gathers rules in an order-independent manner, preventing interference between requests and enabling the removal of original forget-set data post-rule induction. Experimental results show successful suppression of targeted knowledge while maintaining utility, demonstrating scalability for sequential requests. The research paper can be accessed on arXiv.