The Erasure Illusion: Stress-Testing the Generalization of LLM Forgetting Evaluation

Current "forgetting" tests are incomplete: Standard unlearning evaluations only check if models forget exact sentences from target datasets, but this creates an "erasure illusion"—models can pass these tests while still retaining the underlying knowledge in slightly different forms, meaning the core information hasn't actually been removed.

Semantic similarity poses a real problem: When researchers ask models about content that is semantically similar but not identical to what was supposed to be "unlearned," the models often still perform well, suggesting that erasing specific text doesn't erase what the model learned from that text. This is particularly concerning for copyright and safety applications where the goal is to remove knowledge, not just specific sentences.

A new testing framework is needed: The paper proposes Proximal Surrogate Generation (PSG), which automatically creates test cases that are semantically related to the original material but worded differently. This stress-tests whether unlearning actually worked by checking if models still know the underlying concepts.

The problem affects multiple major LLMs: Testing across Llama-3-8B, Qwen2.5-7B, and Zephyr-7B-β shows this isn't an isolated issue with one model family—the gap between passing standard tests and actually forgetting knowledge appears to be a widespread problem in current LLM unlearning approaches.

Implications for responsible AI compliance: If models can appear to comply with unlearning requirements (for copyright or safety reasons) while actually retaining the knowledge, this suggests current regulatory approaches may be insufficient and need more rigorous evaluation standards.