Hidden Costs of Counterfactual Knowledge Training in LLM Unlearning: Two Pitfalls Identified

arXiv·medium signal

Ye et al. find that counterfactual tuning for LLM unlearning — training models to generate fictitious knowledge in place of undesired content — has two previously overlooked pitfalls: knowledge conflict (inconsistencies in counterfactual corpora create conflicting gradients that disrupt parameters) and a second mechanism where the training itself leaks information about what was unlearned. These findings complicate the most promising paradigm for selective knowledge removal.