#logit analysis

Researchers introduce Contrastive Decoding Diffing (CDD), a method that recovers verbatim fine-tuning data from large language models using only logit differences between base and finetuned models, without requiring weight or activation access. This technique significantly improves over prior white-box methods, achieving high verbatim recovery scores on the SDF benchmark, and has important implications for ML safety and interpretability by enabling data leakage detection without full model access. CDD uses a single default configuration across model families ranging from 1B to 32B parameters, and on 19 out of 20 model pairs it achieves a verbatim recovery score of 4 or higher out of 5, whereas the prior Activation Difference Lens (ADL) method never exceeds 3 out of 5 despite requiring full weight access.