Existing Large Language Model Unlearning Evaluations Are Inconclusive
This work addresses a critical methodological gap for researchers and practitioners in AI safety and privacy, though it is incremental as it focuses on improving evaluation practices rather than proposing new unlearning methods.
The paper tackles the problem of unreliable evaluations for machine unlearning in large language models, showing that current protocols can overstate or understate success due to issues like information injection and task variability.
Machine unlearning aims to remove sensitive or undesired data from large language models. However, recent studies suggest that unlearning is often shallow, claiming that removed knowledge can easily be recovered. In this work, we critically examine standard unlearning evaluation practices and uncover key limitations that shake our trust in those findings. First, we show that some evaluations introduce substantial new information into the model, potentially masking true unlearning performance by re-teaching the model during testing. Second, we demonstrate that evaluation outcomes vary significantly across tasks, undermining the generalizability of current evaluation routines. Finally, we find that many evaluations rely on spurious correlations, making their results difficult to trust and interpret. Taken together, these issues suggest that current evaluation protocols may both overstate and understate unlearning success. To address this, we propose two principles for future unlearning evaluations: minimal information injection and downstream task awareness. We validate these principles through a series of targeted experiments, showing how violations of each can lead to misleading conclusions.