Guardado en:
Detalles Bibliográficos
Autores principales: Liu, Jinhan, Yang, Yibo, Lu, Ruiying, Piekos, Piotr, Chen, Yimeng, Wang, Peng, Guo, Dandan
Formato: Preprint
Publicado: 2026
Materias:
Acceso en línea:https://arxiv.org/abs/2601.06827
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866908757195227136
author Liu, Jinhan
Yang, Yibo
Lu, Ruiying
Piekos, Piotr
Chen, Yimeng
Wang, Peng
Guo, Dandan
author_facet Liu, Jinhan
Yang, Yibo
Lu, Ruiying
Piekos, Piotr
Chen, Yimeng
Wang, Peng
Guo, Dandan
contents Detecting pre-training data in Large Language Models (LLMs) is crucial for auditing data privacy and copyright compliance, yet it remains challenging in black-box, zero-shot settings where computational resources and training data are scarce. While existing likelihood-based methods have shown promise, they typically aggregate token-level scores using uniform weights, thereby neglecting the inherent information-theoretic dynamics of autoregressive generation. In this paper, we hypothesize and empirically validate that memorization signals are heavily skewed towards the high-entropy initial tokens, where model uncertainty is highest, and decay as context accumulates. To leverage this linguistic property, we introduce Positional Decay Reweighting (PDR), a training-free and plug-and-play framework. PDR explicitly reweights token-level scores to amplify distinct signals from early positions while suppressing noise from later ones. Extensive experiments show that PDR acts as a robust prior and can usually enhance a wide range of advanced methods across multiple benchmarks.
format Preprint
id arxiv_https___arxiv_org_abs_2601_06827
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle PDR: A Plug-and-Play Positional Decay Framework for LLM Pre-training Data Detection
Liu, Jinhan
Yang, Yibo
Lu, Ruiying
Piekos, Piotr
Chen, Yimeng
Wang, Peng
Guo, Dandan
Computation and Language
Detecting pre-training data in Large Language Models (LLMs) is crucial for auditing data privacy and copyright compliance, yet it remains challenging in black-box, zero-shot settings where computational resources and training data are scarce. While existing likelihood-based methods have shown promise, they typically aggregate token-level scores using uniform weights, thereby neglecting the inherent information-theoretic dynamics of autoregressive generation. In this paper, we hypothesize and empirically validate that memorization signals are heavily skewed towards the high-entropy initial tokens, where model uncertainty is highest, and decay as context accumulates. To leverage this linguistic property, we introduce Positional Decay Reweighting (PDR), a training-free and plug-and-play framework. PDR explicitly reweights token-level scores to amplify distinct signals from early positions while suppressing noise from later ones. Extensive experiments show that PDR acts as a robust prior and can usually enhance a wide range of advanced methods across multiple benchmarks.
title PDR: A Plug-and-Play Positional Decay Framework for LLM Pre-training Data Detection
topic Computation and Language
url https://arxiv.org/abs/2601.06827