Measuring the Depth of LLM Unlearning via Activation Patching

Researchers introduced Unlearning Depth Score (UDS), a white-box metric for measuring how completely large language models forget removed knowledge.

UDS traces residual information through internal activations using activation patching, aiming to reveal whether erased knowledge still lingers in the model.

Across 150 unlearned models and 20 metrics, UDS was the most faithful and robust measure of unlearning depth.

The authors also released code, data, and evaluation guidance to support safer privacy, security, and audit checks for LLM unlearning.