Large language models can reproduce training data, but memorization evaluations mostly measure whether models can be made to do so, not whether they do so without adversarial prompting. We introduce PropMe, a framework that pairs a capability evaluation (prefix attacks) with non-adversarial propensity evaluation, and a propensity metric to relate the two. We check every generation against the model's full indexed training corpus with SimpleTrace, an open-source batch implementation of OLMoTrace span matching. We evaluate Comma, its continually pre-trained descendant DFM Decoder, and OLMo3 32B on Common Pile, Dynaword, and Dolma3, with four non-adversarial settings ranging from no prompt to sentences from unseen in-domain sources. Our results show that natural sentence prompts reproduce far less training text than prefix attacks. Generation with no prompt or a single common word, however, reproduces about as much as prefix attacks, or more, in three out of four model-corpus pairs; most of this text is duplicated licenses, code, and templates. In our setting, continual pre-training reduces long verbatim reproduction of the original corpus only modestly. More broadly, we argue that by testing both capability and propensity, we can attain a more comprehensive picture of a model's memorization behavior.
Unintended Memorization of Sensitive Information in Fine-Tuned Language Models
TUM University Hospital(慕尼黑工业大学医院);Technical University of Munich (TUM)(慕尼黑技术大学);Munich Center for Machine Learning (MCML)(慕尼黑机器学习中心);Department of Computing, Imperial College London(伦敦帝国学院计算系)
Copyright Detective: A Forensic System to Evidence LLMs Flickering Copyright Leakage Risks
Pine AI;The University of Texas at Austin(德克萨斯大学奥斯汀分校);University of Illinois Urbana-Champaign(伊利诺伊大学厄巴纳-香槟分校);Duke University(杜克大学);University of Michigan(密歇根大学);Stevens Institute of Technology(史蒂文斯理工学院);Johns Hopkins University(约翰霍普金斯大学);University of Notre Dame(圣母大学);The Pennsylvania State University(宾夕法尼亚州立大学)
Emergent LLM behaviors are observationally equivalent to data leakage
Christopher Barrie Assistant Professor of Sociology, New York University(纽约大学社会学助理教授);New York University(纽约大学);Petter Törnberg Institute for Logic, Language, and Computation, University of Amsterdam(阿姆斯特丹大学逻辑、语言和计算研究所);University of Amsterdam(阿姆斯特丹大学)
Simple Yet Effective: Extracting Private Data Across Clients in Federated Fine-Tuning of Large Language Models
Harbin Institute of Technology(哈尔滨工业大学);Kuaishou Technology(快手科技);Meta AI;Monash University(墨尔本大学);Fudan University(复旦大学);Shanghai Academy of Artificial Intelligence for Science(上海人工智能科学研究院)
From Anchors to Supervision: Memory-Graph Guided Corpus-Free Unlearning for Large Language Models
College of Computer Science and Artificial Intelligence, Fudan University(复旦大学计算机科学与人工智能学院);Shanghai University of Electric Power(上海电力大学);School of Information Science and Engineering, East China University of Science and Technology(东华大学信息科学与工程学院)
Ethicist: Targeted Training Data Extraction Through Loss Smoothed Soft Prompting and Calibrated Confidence Estimation
Institute for Artificial Intelligence(人工智能研究院);Beijing National Research Center for Information Science and Technology(北京信息科学与技术国家研究中心);Tsinghua University(清华大学)
Teach2Eval: An Indirect Evaluation Method for LLM by Judging How It Teaches
School of Computer Science, Fudan University(复旦大学计算机科学学院);Shanghai Innovation Institute(上海创新研究院);Computer Science Department, NYU Shanghai(纽约大学上海学院);DataGrand Inc.(DataGrand公司)
MUSE: Machine Unlearning Six-Way Evaluation for Language Models
University of Washington(华盛顿大学);Princeton University(普林斯顿大学);University of Southern California(南加州大学);University of Chicago(芝加哥大学);Google Research(谷歌研究院)