Researchers argue the term 'machine unlearning' is overused in LLM work and propose stricter definitions

A new position paper on arXiv argues that the term 'machine unlearning' is being overused in large language model (LLM) research and calls for stricter, more precise definitions. The authors—Sangyeon Yoon, Yeachan Jun, and Albert No—contend that 'machine unlearning' should be reserved for dataset-defined deletion: removing the training influence of a precisely specified 'forget set' so that the resulting model is approximately indistinguishable from a model retrained without that data.

The authors identify several LLM tasks currently labeled as 'unlearning'—such as refusal for harmful requests, entity or knowledge removal, and targeted suppression—that pursue different, often policy-dependent objectives. They argue these tasks require distinct terminology and evaluation baselines, such as alignment, suppression, editing, or obfuscation, rather than sharing the 'unlearning' label.

The paper warns that this terminological confusion is not merely cosmetic. Because papers make different implicit guarantees under the same label, metrics and benchmarks are frequently reused outside their intended scope. This can reward surface-level performance—such as low ROUGE or forget accuracy—even when retraining-equivalence is not tested and derived capabilities remain unexamined.

The authors conclude by calling for stricter terminology tied to explicit guarantees and reference models, and for evaluations that match the claimed objective. They position this as a necessary step to improve rigor and comparability in a field where 'forgetting' is increasingly demanded for regulatory, copyright, and safety reasons.