Large language models have been shown to memorize significant portions of their training data, which they can reproduce when appropriately prompted. This work investigates the impact of simple pruning techniques on this behavior. Our findings reveal that pruning effectively reduces the extent of memorization in LLMs, demonstrating its potential as a foundational approach for mitigating membership inference attacks.
Due to the sensitive nature of personally identifiable information (PII), its owners may have the authority to control its inclusion or request its removal from large-language model (LLM) training. Beyond this, PII may be added or removed from training datasets due to evolving dataset curation techniques, because they were newly scraped for retraining, or because they were included in a new downstream fine-tuning stage. We find that the amount and ease of PII memorization is a dynamic property of a model that evolves throughout training pipelines and depends on commonly altered design choices. We characterize three such novel phenomena: (1) similar-appearing PII seen later in training can elicit memorization of earlier-seen sequences in what we call assisted memorization, and this is a significant factor (in our settings, up to 1/3); (2) adding PII can increase memorization of other PII significantly (in our settings, as much as $\approx\!7.5\times$); and (3) removing PII can lead to other PII being memorized. Model creators should consider these first- and second-order privacy risks when training models to avoid the risk of new PII regurgitation.
Recent copyright agreements between AI companies and content creators underscore the need for fine-grained control over language models' ability to reproduce copyrighted text. Existing defenses-ranging from aggressive unlearning to simplistic output filters-either sacrifice model utility or inadequately address verbatim leakage. We introduce Obliviate, a lightweight post-training method that surgically suppresses exact reproduction of specified sequences while preserving semantic understanding. Obliviate first identifies memorized passages and then, for each target token, minimally adjusts the model's output distribution via a Kullback-Leibler divergence penalty to drive down the probability of exact reproduction. Simultaneously, we enforce a consistency loss on non-target tokens to retain the model's fluency and task performance. We evaluate Obliviate on four popular 6-8B-parameter models (LLaMA-3.1, LLaMA-3.1-Instruct, Qwen-2.5, and Yi-1.5) using synthetic memorization benchmarks and organic copyrighted excerpts (e.g., Moby Dick, Frankenstein, Alice in Wonderland and Les Miserables). Across all settings, Obliviate reduces verbatim recall by two orders of magnitude (e.g., from hundreds of words to fewer than 12) while degrading downstream accuracy by at most 1% on HellaSwag, MMLU, TruthfulQA, and Winogrande. Furthermore, we benchmark Obliviate aganist different unlearning and copyright techniques using the MUSE and CoTaEval benchmarks. These results position Obliviate as a practical, high-fidelity solution for copyright compliance in deployed LLMs.
Long context understanding remains challenging for large language models due to their limited context windows. This paper presents Long Input Fine-Tuning (LIFT), a novel framework for long-context modeling that can improve the long-context performance of arbitrary (short-context) LLMs by dynamically adapting model parameters based on the long input. Importantly, LIFT, rather than endlessly extending the context window size to accommodate increasingly longer inputs in context, chooses to store and absorb the long input in parameter. By fine-tuning the long input into model parameters, LIFT allows short-context LLMs to answer questions even when the required information is not provided in the context during inference. Furthermore, to enhance LIFT performance while maintaining the original in-context learning (ICL) capabilities, we introduce Gated Memory, a specialized attention adapter that automatically balances long input memorization and ICL. We provide a comprehensive analysis of the strengths and limitations of LIFT on long context understanding, offering valuable directions for future research.
In LLM evaluations, reasoning is often distinguished from recall/memorization by performing numerical variations to math-oriented questions. Here we introduce a general variation method for multiple-choice questions that completely dissociates the correct answer from previously seen tokens or concepts, requiring LLMs to understand and reason (rather than memorizing) in order to answer correctly. Using this method, we evaluate state-of-the-art proprietary and open-source LLMs on two datasets available in English and Spanish: the public MMLU benchmark and the private UNED-Access 2024 dataset. Results show that all models experience remarkable accuracy drops under our proposed variation, with an average loss of 57% on MMLU and 50% on UNED-Access 2024, ranging from 10% to 93% across models. Notably, the most accurate model in our experimentation (OpenAI-o3-mini) is not the most robust (DeepSeek-R1-70B), suggesting that the best models in standard evaluations may not be the ones with better reasoning capabilities. Also, we see larger accuracy drops in public (vs private) datasets and questions posed in their original language (vs a manual translation), which are signs of contamination and also point to a relevant role of recall/memorization in current LLMs' answers.
Large Language Models (LLMs) pose significant privacy risks, potentially leaking training data due to implicit memorization. Existing privacy attacks primarily focus on membership inference attacks (MIAs) or data extraction attacks, but reconstructing specific personally identifiable information (PII) in LLM's training data remains challenging. In this paper, we propose R.R. (Recollect and Rank), a novel two-step privacy stealing attack that enables attackers to reconstruct PII entities from scrubbed training data where the PII entities have been masked. In the first stage, we introduce a prompt paradigm named recollection, which instructs the LLM to repeat a masked text but fill in masks. Then we can use PII identifiers to extract recollected PII candidates. In the second stage, we design a new criterion to score each PII candidate and rank them. Motivated by membership inference, we leverage the reference model as a calibration to our criterion. Experiments across three popular PII datasets demonstrate that the R.R. achieves better PII identical performance compared to baselines. These results highlight the vulnerability of LLMs to PII leakage even when training data has been scrubbed. We release the replicate package of R.R. at a link.
Large Language Models (LLMs) demonstrate remarkable capabilities in question answering (QA), but metrics for assessing their reliance on memorization versus retrieval remain underdeveloped. Moreover, while finetuned models are state-of-the-art on closed-domain tasks, general-purpose models like GPT-4o exhibit strong zero-shot performance. This raises questions about the trade-offs between memorization, generalization, and retrieval. In this work, we analyze the extent to which multimodal retrieval-augmented VLMs memorize training data compared to baseline VLMs. Using the WebQA benchmark, we contrast finetuned models with baseline VLMs on multihop retrieval and question answering, examining the impact of finetuning on data memorization. To quantify memorization in end-to-end retrieval and QA systems, we propose several proxy metrics by investigating instances where QA succeeds despite retrieval failing. Our results reveal the extent to which finetuned models rely on memorization. In contrast, retrieval-augmented VLMs have lower memorization scores, at the cost of accuracy (72% vs 52% on WebQA test set). As such, our measures pose a challenge for future work to reconcile memorization and generalization in both Open-Domain QA and joint Retrieval-QA tasks.
Continual learning (CL) is the sub-field of machine learning concerned with accumulating knowledge in dynamic environments. So far, CL research has mainly focused on incremental classification tasks, where models learn to classify new categories while retaining knowledge of previously learned ones. Here, we argue that maintaining such a focus limits both theoretical development and practical applicability of CL methods. Through a detailed analysis of concrete examples - including multi-target classification, robotics with constrained output spaces, learning in continuous task domains, and higher-level concept memorization - we demonstrate how current CL approaches often fail when applied beyond standard classification. We identify three fundamental challenges: (C1) the nature of continuity in learning problems, (C2) the choice of appropriate spaces and metrics for measuring similarity, and (C3) the role of learning objectives beyond classification. For each challenge, we provide specific recommendations to help move the field forward, including formalizing temporal dynamics through distribution processes, developing principled approaches for continuous task spaces, and incorporating density estimation and generative objectives. In so doing, this position paper aims to broaden the scope of CL research while strengthening its theoretical foundations, making it more applicable to real-world problems.
Recent theoretical studies (Kou et al., 2023; Cao et al., 2022) have revealed a sharp phase transition from benign to harmful overfitting when the noise-to-feature ratio exceeds a threshold-a situation common in long-tailed data distributions where atypical data is prevalent. However, harmful overfitting rarely happens in overparameterized neural networks. Further experimental results suggested that memorization is necessary for achieving near-optimal generalization error in long-tailed data distributions (Feldman & Zhang, 2020). We argue that this discrepancy between theoretical predictions and empirical observations arises because previous feature-noise data models overlook the heterogeneous nature of noise across different data classes. In this paper, we refine the feature-noise data model by incorporating class-dependent heterogeneous noise and re-examine the overfitting phenomenon in neural networks. Through a comprehensive analysis of the training dynamics, we establish test loss bounds for the refined model. Our findings reveal that neural networks can leverage "data noise", previously deemed harmful, to learn implicit features that improve the classification accuracy for long-tailed data. Experimental validation on both synthetic and real-world datasets supports our theoretical results.
The memorization capacity of neural networks with a given architecture has been thoroughly studied in many works. Specifically, it is well-known that memorizing $N$ samples can be done using a network of constant width, independent of $N$. However, the required constructions are often quite delicate. In this paper, we consider the natural question of how well feedforward ReLU neural networks can memorize robustly, namely while being able to withstand adversarial perturbations of a given radius. We establish both upper and lower bounds on the possible radius for general $l_p$ norms, implying (among other things) that width logarithmic in the number of input samples is necessary and sufficient to achieve robust memorization (with robustness radius independent of $N$).
Extreme multi-label classification (XMC) seeks to find relevant labels from an extremely large label collection for a given text input. To tackle such a vast label space, current state-of-the-art methods fall into two categories. The one-versus-all (OVA) method uses learnable label embeddings for each label, excelling at memorization (i.e., capturing detailed training signals for accurate head label prediction). In contrast, the dual-encoder (DE) model maps input and label text into a shared embedding space for better generalization (i.e., the capability of predicting tail labels with limited training data), but may fall short at memorization. To achieve generalization and memorization, existing XMC methods often combine DE and OVA models, which involves complex training pipelines. Inspired by the success of retrieval-augmented language models, we propose the Retrieval-augmented Encoders for XMC (RAEXMC), a novel framework that equips a DE model with retrieval-augmented capability for efficient memorization without additional trainable parameter. During training, RAEXMC is optimized by the contrastive loss over a knowledge memory that consists of both input instances and labels. During inference, given a test input, RAEXMC retrieves the top-$K$ keys from the knowledge memory, and aggregates the corresponding values as the prediction scores. We showcase the effectiveness and efficiency of RAEXMC on four public LF-XMC benchmarks. RAEXMC not only advances the state-of-the-art (SOTA) DE method DEXML, but also achieves more than 10x speedup on the largest LF-AmazonTitles-1.3M dataset under the same 8 A100 GPUs training environments.
The evolution of microscopy, beginning with its invention in the late 16th century, has continuously enhanced our ability to explore and understand the microscopic world, enabling increasingly detailed observations of structures and phenomena. In parallel, the rise of data-driven science has underscored the need for sophisticated methods to explore and understand the composition of complex data collections. This paper introduces the Vendiscope, the first algorithmic microscope designed to extend traditional microscopy to computational analysis. The Vendiscope leverages the Vendi scores -- a family of differentiable diversity metrics rooted in ecology and quantum mechanics -- and assigns weights to data points based on their contribution to the overall diversity of the collection. These weights enable high-resolution data analysis at scale. We demonstrate this across biology, materials science, and machine learning (ML). We analyzed the $250$ million protein sequences in the protein universe, discovering that over $200$ million are near-duplicates and that AlphaFold fails on proteins with Gene Ontology (GO) functions that contribute most to diversity. Applying the Vendiscope to the Materials Project database led to similar findings: more than $85\%$ of the crystals with formation energy data are near-duplicates and ML models perform poorly on materials that enhance diversity. Additionally, the Vendiscope can be used to study phenomena such as memorization in generative models. We used the Vendiscope to identify memorized training samples from $13$ different generative models and found that the best-performing ones often memorize the training samples that contribute least to diversity. Our findings demonstrate that the Vendiscope can serve as a powerful tool for data-driven science.
Multi-modal models, such as CLIP, have demonstrated strong performance in aligning visual and textual representations, excelling in tasks like image retrieval and zero-shot classification. Despite this success, the mechanisms by which these models utilize training data, particularly the role of memorization, remain unclear. In uni-modal models, both supervised and self-supervised, memorization has been shown to be essential for generalization. However, it is not well understood how these findings would apply to CLIP, which incorporates elements from both supervised learning via captions that provide a supervisory signal similar to labels, and from self-supervised learning via the contrastive objective. To bridge this gap in understanding, we propose a formal definition of memorization in CLIP (CLIPMem) and use it to quantify memorization in CLIP models. Our results indicate that CLIP's memorization behavior falls between the supervised and self-supervised paradigms, with "mis-captioned" samples exhibiting highest levels of memorization. Additionally, we find that the text encoder contributes more to memorization than the image encoder, suggesting that mitigation strategies should focus on the text domain. Building on these insights, we propose multiple strategies to reduce memorization while at the same time improving utility--something that had not been shown before for traditional learning paradigms where reducing memorization typically results in utility decrease.
Large language models have demonstrated impressive performance on challenging mathematical reasoning tasks, which has triggered the discussion of whether the performance is achieved by true reasoning capability or memorization. To investigate this question, prior work has constructed mathematical benchmarks when questions undergo simple perturbations -- modifications that still preserve the underlying reasoning patterns of the solutions. However, no work has explored hard perturbations, which fundamentally change the nature of the problem so that the original solution steps do not apply. To bridge the gap, we construct MATH-P-Simple and MATH-P-Hard via simple perturbation and hard perturbation, respectively. Each consists of 279 perturbed math problems derived from level-5 (hardest) problems in the MATH dataset (Hendrycksmath et. al., 2021). We observe significant performance drops on MATH-P-Hard across various models, including o1-mini (-16.49%) and gemini-2.0-flash-thinking (-12.9%). We also raise concerns about a novel form of memorization where models blindly apply learned problem-solving skills without assessing their applicability to modified contexts. This issue is amplified when using original problems for in-context learning. We call for research efforts to address this challenge, which is critical for developing more robust and reliable reasoning models.
Deep learning model effectiveness in classification tasks is often challenged by the quality and quantity of training data which, whenever containing strong spurious correlations between specific attributes and target labels, can result in unrecoverable biases in model predictions. Tackling these biases is crucial in improving model generalization and trust, especially in real-world scenarios. This paper presents Diffusing DeBias (DDB), a novel approach acting as a plug-in for common methods in model debiasing while exploiting the inherent bias-learning tendency of diffusion models. Our approach leverages conditional diffusion models to generate synthetic bias-aligned images, used to train a bias amplifier model, to be further employed as an auxiliary method in different unsupervised debiasing approaches. Our proposed method, which also tackles the common issue of training set memorization typical of this type of tech- niques, beats current state-of-the-art in multiple benchmark datasets by significant margins, demonstrating its potential as a versatile and effective tool for tackling dataset bias in deep learning applications.
Diffusion models, known for their tremendous ability to generate high-quality samples, have recently raised concerns due to their data memorization behavior, which poses privacy risks. Recent methods for memory mitigation have primarily addressed the issue within the context of the text modality in cross-modal generation tasks, restricting their applicability to specific conditions. In this paper, we propose a novel method for diffusion models from the perspective of visual modality, which is more generic and fundamental for mitigating memorization. Directly exposing visual data to the model increases memorization risk, so we design a framework where models learn through proxy model parameters instead. Specially, the training dataset is divided into multiple shards, with each shard training a proxy model, then aggregated to form the final model. Additionally, practical analysis of training losses illustrates that the losses for easily memorable images tend to be obviously lower. Thus, we skip the samples with abnormally low loss values from the current mini-batch to avoid memorizing. However, balancing the need to skip memorization-prone samples while maintaining sufficient training data for high-quality image generation presents a key challenge. Thus, we propose IET-AGC+, which redistributes highly memorizable samples between shards, to mitigate these samples from over-skipping. Furthermore, we dynamically augment samples based on their loss values to further reduce memorization. Extensive experiments and analysis on four datasets show that our method successfully reduces memory capacity while maintaining performance. Moreover, we fine-tune the pre-trained diffusion models, e.g., Stable Diffusion, and decrease the memorization score by 46.7\%, demonstrating the effectiveness of our method. Code is available in: https://github.com/liuxiao-guan/IET_AGC.
In a systematic way, we investigate a widely asked question: Do LLMs really understand what they say?, which relates to the more familiar term Stochastic Parrot. To this end, we propose a summative assessment over a carefully designed physical concept understanding task, PhysiCo. Our task alleviates the memorization issue via the usage of grid-format inputs that abstractly describe physical phenomena. The grids represents varying levels of understanding, from the core phenomenon, application examples to analogies to other abstract patterns in the grid world. A comprehensive study on our task demonstrates: (1) state-of-the-art LLMs, including GPT-4o, o1 and Gemini 2.0 flash thinking, lag behind humans by ~40%; (2) the stochastic parrot phenomenon is present in LLMs, as they fail on our grid task but can describe and recognize the same concepts well in natural language; (3) our task challenges the LLMs due to intrinsic difficulties rather than the unfamiliar grid format, as in-context learning and fine-tuning on same formatted data added little to their performance.
Training and fine-tuning large language models (LLMs) with hundreds of billions to trillions of parameters requires tens of thousands of GPUs, and a highly scalable software stack. In this work, we present a novel four-dimensional hybrid parallel algorithm implemented in a highly scalable, portable, open-source framework called AxoNN. We describe several performance optimizations in AxoNN to improve matrix multiply kernel performance, overlap non-blocking collectives with computation, and performance modeling to choose performance optimal configurations. These have resulted in unprecedented scaling and peak flop/s (bf16) for training of GPT-style transformer models on Perlmutter (620.1 Petaflop/s), Frontier (1.381 Exaflop/s) and Alps (1.423 Exaflop/s). While the abilities of LLMs improve with the number of trainable parameters, so do privacy and copyright risks caused by memorization of training data, which can cause disclosure of sensitive or private information at inference time. We highlight this side effect of scale through experiments that explore "catastrophic memorization", where models are sufficiently large to memorize training data in a single pass, and present an approach to prevent it. As part of this study, we demonstrate fine-tuning of a 405-billion parameter LLM using AxoNN on Frontier.
Generative models, particularly text-to-image (T2I) diffusion models, play a crucial role in medical image analysis. However, these models are prone to training data memorization, posing significant risks to patient privacy. Synthetic chest X-ray generation is one of the most common applications in medical image analysis with the MIMIC-CXR dataset serving as the primary data repository for this task. This study presents the first systematic attempt to identify prompts and text tokens in MIMIC-CXR that contribute the most to training data memorization. Our analysis reveals two unexpected findings: (1) prompts containing traces of de-identification procedures (markers introduced to hide Protected Health Information) are the most memorized, and (2) among all tokens, de-identification markers contribute the most towards memorization. This highlights a broader issue with the standard anonymization practices and T2I synthesis with MIMIC-CXR. To exacerbate, existing inference-time memorization mitigation strategies are ineffective and fail to sufficiently reduce the model's reliance on memorized text tokens. On this front, we propose actionable strategies for different stakeholders to enhance privacy and improve the reliability of generative models in medical imaging. Finally, our results provide a foundation for future work on developing and benchmarking memorization mitigation techniques for synthetic chest X-ray generation using the MIMIC-CXR dataset. The anonymized code is available at https://anonymous.4open.science/r/diffusion_memorization-8011/
Large Language Models for Code (LLMs4Code) excel at code generation tasks, yielding promise to release developers from huge software development burdens. Nonetheless, these models have been shown to suffer from the significant privacy risks due to the potential leakage of sensitive information embedded during training, known as the memorization problem. Addressing this issue is crucial for ensuring privacy compliance and upholding user trust, but till now there is a dearth of dedicated studies in the literature that focus on this specific direction. Recently, machine unlearning has emerged as a promising solution by enabling models to "forget" sensitive information without full retraining, offering an efficient and scalable approach compared to traditional data cleaning methods. In this paper, we empirically evaluate the effectiveness of unlearning techniques for addressing privacy concerns in LLMs4Code.Specifically, we investigate three state-of-the-art unlearning algorithms and three well-known open-sourced LLMs4Code, on a benchmark that takes into consideration both the privacy data to be forgotten as well as the code generation capabilites of these models. Results show that it is feasible to mitigate the privacy concerns of LLMs4Code through machine unlearning while maintain their code generation capabilities at the same time. We also dissect the forms of privacy protection/leakage after unlearning and observe that there is a shift from direct leakage to indirect leakage, which underscores the need for future studies addressing this risk.