Distribution Preference Optimization Advances LLM Unlearning

The new DiPO algorithm improves LLM unlearning, achieving top forget quality on the TOFU benchmark while keeping strong performance on MUSE. Read more: getnews.me/distribution-preference-... #llmunlearning #distributionpreference

Simpler Optimizers Boost Robustness in Large Language Model Unlearning

Gradient‑free optimizers improve LLM unlearning robustness; a hybrid using both first‑ and zeroth‑order steps performed best on MUSE and WMDP benchmarks. Read more: getnews.me/simpler-optimizers-boost... #llmunlearning #optimizers

Study Shows LLM Unlearning May Be Reversible, Not True Deletion

A new study finds LLM unlearning often hides data, letting original behavior return after brief fine-tuning; the preprint appeared in May 2025 and was updated in September 2025. getnews.me/study-shows-llm-unlearni... #llmunlearning #machinelearning

StableUN Improves LLM Unlearning by Avoiding Sharp Minima

StableUN targets flatter LLM parameter regions to boost resistance to relearning attacks. Tests on WMDP and MUSE benchmarks showed stronger robustness, reported 24 Sep 2025. getnews.me/stableun-improves-llm-un... #stableun #llmunlearning