Positional Fragility in LLMs: How Offset Effects Reshape Our Understanding of Memorization Risks

• URL: http://arxiv.org/abs/2505.13171v2
• Date: Wed, 28 May 2025 15:39:49 GMT
• Title: Positional Fragility in LLMs: How Offset Effects Reshape Our Understanding of Memorization Risks
• Authors: Yixuan Xu, Antoni-Joan Solergibert i Llaquet, Antoine Bosselut, Imanol Schlag,
• Abstract summary: We show that verbatim memorization is most strongly triggered by short prefixes drawn from the beginning of the context window.<n>We show that shifting sensitive data deeper into the context window suppresses both extractable memorization and degeneration.
• Score: 22.636913794637437
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large language models are known to memorize parts of their training data, posing risk of copyright violations. To systematically examine this risk, we pretrain language models (1B/3B/8B) from scratch on 83B tokens, mixing web-scale data with public domain books used to simulate copyrighted content at controlled frequencies at lengths at least ten times longer than prior work. We thereby identified the offset effect, a phenomenon characterized by two key findings: (1) verbatim memorization is most strongly triggered by short prefixes drawn from the beginning of the context window, with memorization decreasing counterintuitively as prefix length increases; and (2) a sharp decline in verbatim recall when prefix begins offset from the initial tokens of the context window. We attribute this to positional fragility: models rely disproportionately on the earliest tokens in their context window as retrieval anchors, making them sensitive to even slight shifts. We further observe that when the model fails to retrieve memorized content, it often produces degenerated text. Leveraging these findings, we show that shifting sensitive data deeper into the context window suppresses both extractable memorization and degeneration. Our results suggest that positional offset is a critical and previously overlooked axis for evaluating memorization risks, since prior work implicitly assumed uniformity by probing only from the beginning of training sequences.

Related papers

• Finding Dori: Memorization in Text-to-Image Diffusion Models Is Less Local Than Assumed [26.985672667560717]
  Concerns about data privacy and intellectual property remain due to their potential to inadvertently replicate training data.<n>Recent mitigation efforts have focused on identifying and pruning weights responsible for triggering replication.<n>We demonstrate that even after pruning, minor adjustments to text embeddings of input prompts are sufficient to re-trigger data replication.
  arXiv  Detail & Related papers  (2025-07-22T15:02:38Z)
• ExpShield: Safeguarding Web Text from Unauthorized Crawling and Language Modeling Exploitation [17.71790411163849]
  We propose ExpShiled, a proactive self-defense mechanism that mitigates sample-specific memorization via imperceptible text perturbations.<n>Our approach requires no external collaboration while maintaining original readability.<n>Even with privacy backdoors, the Membership Inference Attack (MIA) AUC drops from 0.95 to 0.55, and instance exploitation approaches zero.
  arXiv  Detail & Related papers  (2024-12-30T17:52:02Z)
• Unlocking Memorization in Large Language Models with Dynamic Soft Prompting [66.54460367290146]
  Large language models (LLMs) have revolutionized natural language processing (NLP) tasks such as summarization, question answering, and translation. LLMs pose significant security risks due to their tendency to memorize training data, leading to potential privacy breaches and copyright infringement. We propose a novel method for estimating LLM memorization using dynamic, prefix-dependent soft prompts.
  arXiv  Detail & Related papers  (2024-09-20T18:56:32Z)
• Detecting, Explaining, and Mitigating Memorization in Diffusion Models [49.438362005962375]
  We introduce a straightforward yet effective method for detecting memorized prompts by inspecting the magnitude of text-conditional predictions. Our proposed method seamlessly integrates without disrupting sampling algorithms, and delivers high accuracy even at the first generation step. Building on our detection strategy, we unveil an explainable approach that shows the contribution of individual words or tokens to memorization.
  arXiv  Detail & Related papers  (2024-07-31T16:13:29Z)
• Demystifying Verbatim Memorization in Large Language Models [67.49068128909349]
  Large Language Models (LLMs) frequently memorize long sequences verbatim, often with serious legal and privacy implications. We develop a framework to study verbatim memorization in a controlled setting by continuing pre-training from Pythia checkpoints with injected sequences. We find that (1) non-trivial amounts of repetition are necessary for verbatim memorization to happen; (2) later (and presumably better) checkpoints are more likely to memorize verbatim sequences, even for out-of-distribution sequences.
  arXiv  Detail & Related papers  (2024-07-25T07:10:31Z)
• Uncovering Latent Memories: Assessing Data Leakage and Memorization Patterns in Frontier AI Models [7.50189359952191]
  We show that sequences which are not memorized after the first encounter can be "uncovered" throughout the course of training. The presence of latent memorization presents a challenge for data privacy as memorized sequences may be hidden at the final checkpoint of the model. We develop a diagnostic test relying on the cross entropy loss to uncover latent memorized sequences with high accuracy.
  arXiv  Detail & Related papers  (2024-06-20T17:56:17Z)
• Exploring Memorization in Fine-tuned Language Models [53.52403444655213]
  We conduct the first comprehensive analysis to explore language models' memorization during fine-tuning across tasks. Our studies with open-sourced and our own fine-tuned LMs across various tasks indicate that memorization presents a strong disparity among different fine-tuning tasks. We provide an intuitive explanation of this task disparity via sparse coding theory and unveil a strong correlation between memorization and attention score distribution.
  arXiv  Detail & Related papers  (2023-10-10T15:41:26Z)
• Ethicist: Targeted Training Data Extraction Through Loss Smoothed Soft Prompting and Calibrated Confidence Estimation [56.57532238195446]
  We propose a method named Ethicist for targeted training data extraction. To elicit memorization, we tune soft prompt embeddings while keeping the model fixed. We show that Ethicist significantly improves the extraction performance on a recently proposed public benchmark.
  arXiv  Detail & Related papers  (2023-07-10T08:03:41Z)
• Mitigating Approximate Memorization in Language Models via Dissimilarity Learned Policy [0.0]
  Large Language models (LLMs) are trained on large amounts of data. LLMs showed to memorize parts of the training data and emit those data verbatim when an adversary prompts appropriately.
  arXiv  Detail & Related papers  (2023-05-02T15:53:28Z)
• Unintended Memorization and Timing Attacks in Named Entity Recognition Models [5.404816271595691]
  We study the setting when NER models are available as a black-box service for identifying sensitive information in user documents. With updated pre-trained NER models from spaCy, we demonstrate two distinct membership attacks on these models.
  arXiv  Detail & Related papers  (2022-11-04T03:32:16Z)
• Preventing Verbatim Memorization in Language Models Gives a False Sense of Privacy [91.98116450958331]
  We argue that verbatim memorization definitions are too restrictive and fail to capture more subtle forms of memorization. Specifically, we design and implement an efficient defense that perfectly prevents all verbatim memorization. We conclude by discussing potential alternative definitions and why defining memorization is a difficult yet crucial open question for neural language models.
  arXiv  Detail & Related papers  (2022-10-31T17:57:55Z)
• Quantifying Memorization Across Neural Language Models [61.58529162310382]
  Large language models (LMs) have been shown to memorize parts of their training data, and when prompted appropriately, they will emit the memorized data verbatim. This is undesirable because memorization violates privacy (exposing user data), degrades utility (repeated easy-to-memorize text is often low quality), and hurts fairness (some texts are memorized over others). We describe three log-linear relationships that quantify the degree to which LMs emit memorized training data.
  arXiv  Detail & Related papers  (2022-02-15T18:48:31Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.