margin-self-correction

Introduces MASC (Margin Self-Correction), an efficient unlearning method for LLMs that uses an online stopping rule to achieve competitive forget–retain trade-offs at reduced computational cost, validated on TOFU and MUSE benchmarks.