A Comprehensive Evaluation of LLM Unlearning Robustness under Multi-Turn Interaction

Researchers found that machine unlearning in large language models, which aims to remove specific training data influence, is less effective in interactive settings than previously thought. Knowledge that appears forgotten in static tests can often be recovered through multi-turn conversations and self-correction interactions.

• →Machine unlearning effectiveness is overestimated when evaluated only in static, single-turn settings.
• →Knowledge appearing forgotten can be recovered through interactive patterns like self-correction and dialogue-conditioned querying.
• →Stronger unlearning methods often result in behavioral rigidity rather than genuine knowledge erasure.
• →Current evaluation methods may not accurately reflect real-world unlearning robustness.
• →Interactive environments pose significant challenges for maintaining stable knowledge forgetting in LLMs.