Researchers have proposed a new method called JUMP (Joint Uncertainty-Guided Mask Probing) to detect whether individual records were used during the fine-tuning of open-weight language models. Unlike previous methods, JUMP can identify sensitive information in parallel and requires fewer model evaluations. The authors tested JUMP on six different domains and found it improved mean ROC-AUC scores compared to a previous method called SAMA. According to the study, JUMP achieved higher accuracy with only three model forwards per sample, whereas SAMA required 32 forwards. The proposed method is intended for auditing individual records in fine-tuned language models. --- Why it matters: This matters because it provides a more efficient and accurate way to detect whether sensitive information has been used during the fine-tuning of open-weight language models. This can help ensure that such models are being used responsibly and comply with data protection regulations. Source: https://arxiv.org/abs/2607.16207