Research map / Large language models / Language, safety and interpretability
Enhancing Pre-training Data Detection in LLMs Through Discriminative and Symmetric Prefix Selection
Kai Sun, Yuxin Lin, Bo Dong, Jingyao Zhang, Bin Shi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| select prefixes with Discriminability/Symmetry criteria instead of random unlabeled prefixes | RECALL · WikiMIA-32 (avg over 5 LLMs) | AUC: 65.44 -> 82.72 | Improved |
| select prefixes with DSC-Prefix instead of labeled random prefixes | RECALL · WikiMIA-32 (avg) | AUC: 77.60 -> 82.72 | Improved |
| select prefix pairs with DSC-Prefix instead of labeled random | CON-RECALL · WikiMIA-32 (avg) | AUC: 79.30 -> 86.92 | Improved |
| DSC-Prefix vs labeled prefix on WikiMIA-256 | RECALL · WikiMIA-256 (avg) | AUC: 82.24 -> 84.30 | Improved |
| DSC-Prefix with extremely imbalanced non-member:member ratio 1:9 | RECALL · imbalanced WikiMIA | AUC | Hurt |
| DSC-Prefix selection when classes are roughly balanced | RECALL · imbalanced WikiMIA | AUC | Improved |
| select prefixes on 20% subset of test samples (ablation) | DSC-Prefix · WikiMIA | AUC | No clear effect |
| raise quantile proportions (more candidate prefixes) (ablation) | DSC-Prefix · WikiMIA | AUC | Hurt |
| use few-shot (concatenated) prefixes with DSC-Prefix | RECALL · WikiMIA-32/64 Pythia-6.9B | AUC | Improved |
| use member prefixes vs non-member prefixes | RECALL · WikiMIA | AUC | Hurt |
Closest papers by meaning
- Detecting Pretraining Data from Large Language Models (ICLR 2024 · similarity 0.89)
- Improving Pretraining Data Using Perplexity Correlations (ICLR 2025 · similarity 0.88)
- Enhancing Multilingual LLM Pretraining with Model-Based Data Selection (NeurIPS 2025 · similarity 0.87)
- Fine-tuning can Help Detect Pretraining Data from Large Language Models (ICLR 2025 · similarity 0.87)
- MATES: Model-Aware Data Selection for Efficient Pretraining with Data Influence Models (NeurIPS 2024 · similarity 0.86)
- Training-free LLM-generated Text Detection by Mining Token Probability Sequences (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
