Audit of 263 Batch-Normalized Checkpoints Finds Unlearning Numbers Move Per Checkpoint
This matters to developers because it reveals a potential flaw in how unlearning is verified for these models, which could impact trust in AI systems that rely on unlearning capabilities.