Federated large language model unlearning aims to erase specific knowledge from a model when the data to forget and the data to retain are distributed across private clients and never shared with the server. This setup makes it impossible for the server to directly inspect or separate the data, so it must rely on updates sent from clients. Those updates, however, are often locally trained adapters that encode conflicting information, which can undermine the unlearning process.

The new work, presented in an arXiv preprint, introduces FIRM (Fisher-Informed Retention Merging), a method designed to handle these conflicting adapter updates. By using Fisher information, FIRM weighs and merges the retention-focused adapters from different clients, aiming to preserve the knowledge that should be kept while suppressing the targeted knowledge that must be removed.

The abstract notes that the locally trained adapters encode conflicting signals, but the full details of the method and its evaluation are not included in the available text. As a single-source report, this summary reflects only what the abstract states; no comparison with other approaches or experimental results is provided here.