Unlearning in LLMs leaves detectable 'fingerprints' that let a simple classifier distinguish an unlearned model from its original, even on unrelated prompts.
forget” prompts, and above 98% on UltraChat for all models. Train- ing on forget-only data (Sf) likewise yields over 97% detection on “forget irrelevant
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Unlearning Isn't Invisible: Detecting Unlearning Traces in LLMs from Model Outputs
Unlearning in LLMs leaves detectable 'fingerprints' that let a simple classifier distinguish an unlearned model from its original, even on unrelated prompts.