ResearchResearch paperLarge Language Models · Training & Scaling1 source · Oct 12, 2026

WildTrace: Evaluating Training Data Detection in Realistic Large-Scale LLM Post-Training

To this end, we introduce \method, a reliable benchmark designed to evaluate detection methods in realistic, large-scale post-training settings.

Although post-training significantly enhances model performance, the use of massive undisclosed data raises concerns about benchmark contamination and copyright infringement.
This drives growing interest in training data detection for post-trained models.
Existing evaluations for training data detection often leverage benchmarks constructed post hoc, which can introduce distribution shifts.
To perform a reliable evaluation, we construct model-specific evaluation sets across four diverse tasks, featuring distribution-matched members and non-members with verifiable membership labels.

[1]WildTrace: Evaluating Training Data Detection in Realistic Large-Scale LLM Post-Training
arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 04:21 AM
To this end, we introduce \method, a reliable benchmark designed to evaluate detection methods in realistic, large-scale post-training settings.
Although post-training significantly enhances model performance, the use of massive undisclosed data raises concerns about benchmark contamination and copyright infringement.

Extractive summary: sentences quoted from the sources.

Before this