Measuring the Depth of LLM Unlearning via Activation Patching (gnueaj.github.io)

Researchers have introduced the Unlearning Depth Score (UDS), a novel metric designed to quantify the depth of knowledge unlearning in large language models (LLMs) through a mechanism called activation patching. This two-stage process compares the model's hidden states before and after targeted unlearning, allowing the researchers to evaluate how much of the knowledge has been successfully erased. The development includes an interactive UDS pipeline, a comprehensive meta-evaluation of 20 existing metrics, and benchmarking across 150 models utilizing the Open-Unlearning framework with the Llama-3.2-1B-Instruct model. This advancement is significant for the AI/ML community as it enhances the understanding of knowledge retention and removal in LLMs, particularly in the context of privacy and data management. By measuring how effective various unlearning techniques are, the UDS provides a more reliable method to ensure that sensitive information is not recoverable post-unlearning. The results suggest that the UDS outperforms traditional metrics in both faithfulness and robustness, ultimately offering a clearer picture of a model's ability to safeguard privacy while maintaining general capabilities. This reflects a critical shift towards more accountable and ethical AI development.