# Data leakage

5 related articles

Subscribe to updates

arXiv Oct 16, 2026
Pruning 1.6% of neurons cuts name inference in fine-tuned LLMs by up to 7.4 points, with no retraining
Large language models fine-tuned on proprietary records memorize names and other personal data, and existing privacy-neuron editing inspects only the leaking model, with no PII-free reference to separate identity neurons from useful context. JEPA-Prune trains a public model on surrogate and masked views of the same records, never on the original personal values, and uses a cross-model representation probe, integrated gradients and a scale-invariant specificity score to locate PII-responsive neurons, which are zeroed before deployment without retraining. Across three PII datasets and five models, pruning 1.6% of neurons lowers PII inference accuracy by up to 7.4 and extraction recall by up to 13.9, while unseen-text perplexity stays within 16% of the unpruned model for most LLMs.
Large language models fine-tuned on proprietary records memorize names and other personal data, and existing privacy-neuron editing inspects only the leaking model, with no PII-free reference to separate identity neurons from useful context. JEPA-Prune trains a public model on surrogate and masked views of the same records, never on the original personal values, and uses a cross-model representation probe, integrated gradients and a scale-invariant specificity score to locate PII-responsive neurons, which are zeroed before deployment without retraining. Across three PII datasets and five models, pruning 1.6% of neurons lowers PII inference accuracy by up to 7.4 and extraction recall by up to 13.9, while unseen-text perplexity stays within 16% of the unpruned model for most LLMs.
Washington State University
Mohamed Shaaban（First author）
JEPA-Prune Joint-Embedding Predictive Architecture Data leakage Integrated gradients LoRA
arXiv Oct 14, 2026
Data science agents often score well while breaking method rules: across 220 tasks, more data leakage tracked with higher predictive performance
An evaluation across 220 data science tasks finds most agentic solutions contain at least one methodological failure, and that data leakage failures correlate positively with predictive performance, so accuracy alone misses method problems.
Xue Li（First author）
Agentic data science evaluation Failure mode analysis Data leakage Curated data science test suite
researchsquare.com Oct 20, 2026
Nearly half of a widely used diabetes symptom dataset repeats itself: grouping by profile drops ensemble accuracy from 98% to 90–93%
In the UCI Early-Stage Diabetes dataset, 269 of 520 questionnaire records repeat another record's full predictor vector, and row-level cross-validation splits copies of one profile across training and test folds. Grouping all records of a profile into the same fold lowered the Matthews correlation coefficient of ten ensembles by 0.105 to 0.157 and accuracy from 97.2–98.2% to 90.2–93.2%, and changed model rankings. Grouping near-duplicate symptom patterns that differ only in age lowered performance further for every model.
In the UCI Early-Stage Diabetes dataset, 269 of 520 questionnaire records repeat another record's full predictor vector, and row-level cross-validation splits copies of one profile across training and test folds. Grouping all records of a profile into the same fold lowered the Matthews correlation coefficient of ten ensembles by 0.105 to 0.157 and accuracy from 97.2–98.2% to 90.2–93.2%, and changed model rankings. Grouping near-duplicate symptom patterns that differ only in age lowered performance further for every model.
Universidad Nacional José Faustino Sánchez Carrión
William Marin-Rodriguez（First author）
Data leakage Profile-Grouped Cross-Validation Early-Stage Diabetes Risk Prediction Dataset Ensemble Learning
arXiv Oct 12, 2026
Perovskite model scoring gets a new split: a lookup table matches the published best on random splits, and neural models overtake trees on unseen institutions
Evaluating perovskite solar-cell machine learning with provenance-grouped splits shows a lookup table that learns nothing matches the published state of the art on a random split, while neural models overtake tree ensembles on unseen institutions and later years.
Duke University Georgia Institute of Technology
Qiwei Han（First author）
Provenance-grouped protocol Perovskite Database Data leakage Benchmarking thirteen models perovskite solar cells
arXiv Oct 5, 2026
Whiteout overwrites personal sensitive information with targeted obfuscation samples, preventing LLMs from regurgitating it while leaving utility and safety largely intact
The work presents Whiteout, a practical tool that, upon requests by individuals, prevents LLMs from regurgitating their genuine personally sensitive information (PSI) by overwriting it with precise and carefully designed obfuscation samples; evaluated on modern LLMs of varying sizes and makers, including a widely-used OpenAI model, Whiteout effectively prevents disclosure of the targeted PSIs, has negligible impact on model utility and safety, outperforms existing alternatives, and is tested against countermeasures ranging from black-box attacks such as jailbreaking to white-box adaptive attacks such as relearning and quantization.
The work presents Whiteout, a practical tool that, upon requests by individuals, prevents LLMs from regurgitating their genuine personally sensitive information (PSI) by overwriting it with precise and carefully designed obfuscation samples; evaluated on modern LLMs of varying sizes and makers, including a widely-used OpenAI model, Whiteout effectively prevents disclosure of the targeted PSIs, has negligible impact on model utility and safety, outperforms existing alternatives, and is tested against countermeasures ranging from black-box attacks such as jailbreaking to white-box adaptive attacks such as relearning and quantization.
University of Chicago
Anna Yoo Jeong Ha（First author）
Whiteout Data leakage Large language models

You’re all caught up

Pruning 1.6% of neurons cuts name inference in fine-tuned LLMs by up to 7.4 points, with no retraining

Data science agents often score well while breaking method rules: across 220 tasks, more data leakage tracked with higher predictive performance

Nearly half of a widely used diabetes symptom dataset repeats itself: grouping by profile drops ensemble accuracy from 98% to 90–93%

Perovskite model scoring gets a new split: a lookup table matches the published best on random splits, and neural models overtake trees on unseen institutions

Whiteout overwrites personal sensitive information with targeted obfuscation samples, preventing LLMs from regurgitating it while leaving utility and safety largely intact