Tag
The paper introduces Diff Mining, a framework for identifying finetuning objectives in language models by analyzing logit differences between finetuned and base models, enabling interpretable auditing of learned behaviors.
This paper introduces Contrastive Decoding Diffing (CDD), a method that recovers verbatim finetuning data from LLMs using only logit access, without requiring model weights or activations. It demonstrates high verbatim recovery across multiple model families and reveals unintended data leakage, such as the recurring fictional persona 'Dr. Elena Rodriguez' from synthetic training data.