Researchers just made it cheaper to teach an AI model to forget something.
A new paper introduces Mamushi, a framework for what its authors call distributional unlearning - removing an entire category of data from a trained model's influence, like toxic language or a specific topic, rather than erasing one record at a time. Mamushi ranks which examples to delete using a classifier that estimates the ratio between how likely a piece of data belongs to the unwanted "forget" domain versus the domain you want to keep. The paper proves that setting a threshold on that ratio is the mathematically optimal way to pick a fixed number of examples for removal, and the authors back the theory with tests on real datasets covering toxic-language removal and topic removal.
That matters because most unlearning methods lean on simplifying assumptions about how data is distributed, and those assumptions tend to fall apart in the sprawling, high-dimensional spaces where language-model representations actually live. Mamushi skips those assumptions and, per the paper, reaches the same forgetting target while deleting fewer examples - which means less damage to everything else the model still needs to know. That efficiency gain matters well beyond academia: regulations like the EU's GDPR already give people a right to have their data forgotten, and that obligation gets expensive fast when "forgetting" means touching a multi-billion-parameter model.
A sharper selection rule is a genuine improvement, not a breakthrough - it is one piece of the unlearning pipeline, not proof that a model can ever fully un-know something it was trained on.