Forgetting Only What Matters: Layer-Selective Unlearning toward Robust LLMs

Machine unlearning in large language models often falters after downstream compression, where low-bit post-training quantization can inadvertently revive scrubbed memorization. To address this brittleness, researchers introduce FOM-UL, a layer-selective unlearning framework that ranks transformer layers using a forget-to-retain significance score. By confining parameter updates to layers heavily tied to the target knowledge while sparing the rest of the network, FOM-UL mitigates collateral damage to general utility. Benchmark evaluations across TOFU, KnowUnDo, and MUSE demonstrate that FOM-UL sustains unlearning efficacy and retain-set performance even under 8-bit and 4-bit quantization, curbing leakage against adversarial prompt probing.