Machine Unlearning for LLMs

Why it matters

Right-to-be-forgotten requires unlearning. Understanding shapes compliance.

The architecture

Approaches: fine-tune to unlearn, gradient ascent, targeted forgetting.

How it works end to end

Full retrain: gold standard, expensive.

Approximate: gradient ascent on forget set.

Fine-tune: targeted intervention.

Verification: MIA post-unlearn.