AdaPop Unlearning Method Reduces Information Leakage via Popularity-Aware Gradients
August 13, 2026
AdaPop uses a popularity-dependent exponent and a dual-ascent controller to address uneven memorization in LLMs. The method achieves 5x less leakage of forgotten content under paraphrased queries and 1.6x less under adversarial attacks compared to uniform gradient methods.
HOW THIS AFFECTS YOU
●
researcherYou can better address the disproportionate memorization of high-frequency facts during unlearning.
●
policyThis provides a more robust technical approach for meeting data deletion and privacy compliance requirements.