Gaussian blur is widely used to blur human faces in sensitive photos before the photos are posted on the Internet. However, it is unclear to what extent the blurred faces can be restored and used to re-identify the person, especially under a high-blurring setting. In this paper, we explore this question by developing a deblurring method called Revelio. The key intuition is to leverage a generative model's memorization effect and approximate the inverse function of Gaussian blur for face restoration. Compared with existing methods, we design the deblurring process to be identity-preserving. It uses a conditional Diffusion model for preliminary face restoration and then uses an identity retrieval model to retrieve related images to further enhance fidelity. We evaluate Revelio with large public face image datasets and show that it can effectively restore blurred faces, especially under a high-blurring setting. It has a re-identification accuracy of 95.9%, outperforming existing solutions. The result suggests that Gaussian blur should not be used for face anonymization purposes. We also demonstrate the robustness of this method against mismatched Gaussian kernel sizes and functions, and test preliminary countermeasures and adaptive attacks to inspire future work.
Data reconstruction attacks, which aim to recover the training dataset of a target model with limited access, have gained increasing attention in recent years. However, there is currently no consensus on a formal definition of data reconstruction attacks or appropriate evaluation metrics for measuring their quality. This lack of rigorous definitions and universal metrics has hindered further advancement in this field. In this paper, we address this issue in the vision domain by proposing a unified attack taxonomy and formal definitions of data reconstruction attacks. We first propose a set of quantitative evaluation metrics that consider important criteria such as quantifiability, consistency, precision, and diversity. Additionally, we leverage large language models (LLMs) as a substitute for human judgment, enabling visual evaluation with an emphasis on high-quality reconstructions. Using our proposed taxonomy and metrics, we present a unified framework for systematically evaluating the strengths and limitations of existing attacks and establishing a benchmark for future research. Empirical results, primarily from a memorization perspective, not only validate the effectiveness of our metrics but also offer valuable insights for designing new attacks.
Large Language Models (LLMs) memorize, and thus, among huge amounts of uncontrolled data, may memorize Personally Identifiable Information (PII), which should not be stored and, consequently, not leaked. In this paper, we introduce Private Memorization Editing (PME), an approach for preventing private data leakage that turns an apparent limitation, that is, the LLMs' memorization ability, into a powerful privacy defense strategy. While attacks against LLMs have been performed exploiting previous knowledge regarding their training data, our approach aims to exploit the same kind of knowledge in order to make a model more robust. We detect a memorized PII and then mitigate the memorization of PII by editing a model knowledge of its training data. We verify that our procedure does not affect the underlying language model while making it more robust against privacy Training Data Extraction attacks. We demonstrate that PME can effectively reduce the number of leaked PII in a number of configurations, in some cases even reducing the accuracy of the privacy attacks to zero.
Generative models, with their success in image and video generation, have recently been explored for synthesizing effective neural network weights. These approaches take trained neural network checkpoints as training data, and aim to generate high-performing neural network weights during inference. In this work, we examine four representative methods on their ability to generate novel model weights, i.e., weights that are different from the checkpoints seen during training. Surprisingly, we find that these methods synthesize weights largely by memorization: they produce either replicas, or at best simple interpolations, of the training checkpoints. Current methods fail to outperform simple baselines, such as adding noise to the weights or taking a simple weight ensemble, in obtaining different and simultaneously high-performing models. We further show that this memorization cannot be effectively mitigated by modifying modeling factors commonly associated with memorization in image diffusion models, or applying data augmentations. Our findings provide a realistic assessment of what types of data current generative models can model, and highlight the need for more careful evaluation of generative models in new domains. Our code is available at https://github.com/boyazeng/weight_memorization.
While Large Language Models (LLMs) have demonstrated significant advancements in reasoning and agent-based problem-solving, current evaluation methodologies fail to adequately assess their capabilities: existing benchmarks either rely on closed-ended questions prone to saturation and memorization, or subjective comparisons that lack consistency and rigor. In this work, we introduce HeuriGym, an agentic framework designed for evaluating heuristic algorithms generated by LLMs for combinatorial optimization problems, characterized by clearly defined objectives and expansive solution spaces. HeuriGym empowers LLMs to propose heuristics, receive evaluative feedback via code execution, and iteratively refine their solutions. We evaluate nine state-of-the-art models on nine problems across domains such as computer systems, logistics, and biology, exposing persistent limitations in tool use, planning, and adaptive reasoning. To quantify performance, we propose the Quality-Yield Index (QYI), a metric that captures both solution pass rate and quality. Even top models like GPT-o4-mini-high and Gemini-2.5-Pro attain QYI scores of only 0.6, well below the expert baseline of 1. Our open-source benchmark aims to guide the development of LLMs toward more effective and realistic problem-solving in scientific and engineering domains.
Recent work on diffusion models proposed that they operate in two regimes: memorization, in which models reproduce their training data, and generalization, in which they generate novel samples. While this has been tested in high-noise settings, the behavior of diffusion models as effective denoisers when the corruption level is small remains unclear. To address this gap, we systematically investigated the behavior of diffusion models under low-noise diffusion dynamics, with implications for model robustness and interpretability. Using (i) CelebA subsets of varying sample sizes and (ii) analytic Gaussian mixture benchmarks, we reveal that models trained on disjoint data diverge near the data manifold even when their high-noise outputs converge. We quantify how training set size, data geometry, and model objective choice shape denoising trajectories and affect score accuracy, providing insights into how these models actually learn representations of data distributions. This work starts to address gaps in our understanding of generative model reliability in practical applications where small perturbations are common.
Federated fine-tuning of large language models (FedLLMs) presents a promising approach for achieving strong model performance while preserving data privacy in sensitive domains. However, the inherent memorization ability of LLMs makes them vulnerable to training data extraction attacks. To investigate this risk, we introduce simple yet effective extraction attack algorithms specifically designed for FedLLMs. In contrast to prior "verbatim" extraction attacks, which assume access to fragments from all training data, our approach operates under a more realistic threat model, where the attacker only has access to a single client's data and aims to extract previously unseen personally identifiable information (PII) from other clients. This requires leveraging contextual prefixes held by the attacker to generalize across clients. To evaluate the effectiveness of our approaches, we propose two rigorous metrics-coverage rate and efficiency-and extend a real-world legal dataset with PII annotations aligned with CPIS, GDPR, and CCPA standards, achieving 89.9% human-verified precision. Experimental results show that our method can extract up to 56.57% of victim-exclusive PII, with "Address," "Birthday," and "Name" being the most vulnerable categories. Our findings underscore the pressing need for robust defense strategies and contribute a new benchmark and evaluation framework for future research in privacy-preserving federated learning.
Recently, knowledge editing (KE) has emerged as a promising approach to update specific facts in Large Language Models (LLMs) without the need for full retraining. Despite the effectiveness in general-domain benchmarks, their applicability to complex medical domain remains largely unexplored. Medical knowledge editing is particularly challenging, as it requires LLMs to internalize the knowledge and generalize to unseen scenarios for effective and interpretable decision-making. In this work, we propose a novel framework called MedEditBench to rigorously evaluate the effectiveness of existing KE methods in the medical domain. In MedEditBench, we introduce a new medical knowledge editing benchmark as well as three different knowledge editing paradigms, which are designed to assess the impact of different knowledge sources for editing. Our findings indicate that current KE methods result in only superficial memorization of the injected information, failing to generalize to new scenarios. To overcome this limitation, we present Self-Generated Rationale Editing (SGR-Edit), which utilizes model-derived rationales as the target knowledge for editing, thereby uncovering the underlying reasoning process and demonstrating significant improvements over existing KE approaches. Additionally, we offer deeper insights into medical knowledge editing, including the localization of medical knowledge in LLMs and the impact of sequential editing on evolving knowledge. This could provide practical guidance for implementing KE methods in real-world medical applications.
Understanding what and how neural networks memorize during training is crucial, both from the perspective of unintentional memorization of potentially sensitive information and from the standpoint of effective knowledge acquisition for real-world, knowledge-intensive tasks. While previous studies primarily investigate memorization within a single modality, such as text memorization in large language models or image memorization in diffusion models, unified multimodal models are becoming increasingly prevalent in practical applications. In this work, we focus on the unique characteristics of cross-modality memorization and conduct a systematic study centered on vision-language models. To facilitate controlled experiments, we first introduce a synthetic persona dataset comprising diverse synthetic person images and textual descriptions. We quantify factual knowledge memorization and cross-modal transferability by training models on a single modality and evaluating their performance in the other. Our results reveal that facts learned in one modality transfer to the other, but a significant gap exists between recalling information in the source and target modalities. Furthermore, we observe that this gap exists across various scenarios, including more capable models, machine unlearning, and the multi-hop case. At the end, we propose a baseline method to mitigate this challenge. We hope our study can inspire future research on developing more robust multimodal learning techniques to enhance cross-modal transferability.
Recent research demonstrated that training large language models involves memorization of a significant fraction of training data. Such memorization can lead to privacy violations when training on sensitive user data and thus motivates the study of data memorization's role in learning. In this work, we develop a general approach for proving lower bounds on excess data memorization, that relies on a new connection between strong data processing inequalities and data memorization. We then demonstrate that several simple and natural binary classification problems exhibit a trade-off between the number of samples available to a learning algorithm, and the amount of information about the training data that a learning algorithm needs to memorize to be accurate. In particular, $\Omega(d)$ bits of information about the training data need to be memorized when $O(1)$ $d$-dimensional examples are available, which then decays as the number of examples grows at a problem-specific rate. Further, our lower bounds are generally matched (up to logarithmic factors) by simple learning algorithms. We also extend our lower bounds to more general mixture-of-clusters models. Our definitions and results build on the work of Brown et al. (2021) and address several limitations of the lower bounds in their work.
Sequence models, such as Large Language Models (LLMs) and autoregressive image generators, have a tendency to memorize and inadvertently leak sensitive information. While this tendency has critical legal implications, existing tools are insufficient to audit the resulting risks. We hypothesize that those tools' shortcomings are due to mismatched assumptions. Thus, we argue that effectively measuring privacy leakage in sequence models requires leveraging the correlations inherent in sequential generation. To illustrate this, we adapt a state-of-the-art membership inference attack to explicitly model within-sequence correlations, thereby demonstrating how a strong existing attack can be naturally extended to suit the structure of sequence models. Through a case study, we show that our adaptations consistently improve the effectiveness of memorization audits without introducing additional computational costs. Our work hence serves as an important stepping stone toward reliable memorization audits for large sequence models.
We propose a new method for estimating how much a model ``knows'' about a datapoint and use it to measure the capacity of modern language models. Prior studies of language model memorization have struggled to disentangle memorization from generalization. We formally separate memorization into two components: \textit{unintended memorization}, the information a model contains about a specific dataset, and \textit{generalization}, the information a model contains about the true data-generation process. When we completely eliminate generalization, we can compute the total memorization, which provides an estimate of model capacity: our measurements estimate that GPT-style models have a capacity of approximately 3.6 bits per parameter. We train language models on datasets of increasing size and observe that models memorize until their capacity fills, at which point ``grokking'' begins, and unintended memorization decreases as models begin to generalize. We train hundreds of transformer language models ranging from $500K$ to $1.5B$ parameters and produce a series of scaling laws relating model capacity and data size to membership inference.
In this paper, we aim to establish a simple, effective, and theoretically grounded benchmark for rigorously probing abstract reasoning in Large Language Models (LLMs). To achieve this, we first develop a mathematic framework that defines abstract reasoning as the ability to: (i) extract essential patterns independent of surface representations, and (ii) apply consistent rules to these abstract patterns. Based on this framework, we introduce two novel complementary metrics: \(\scoreGamma\) measures basic reasoning accuracy, while \(\scoreDelta\) quantifies a model's reliance on specific symbols rather than underlying patterns - a key indicator of true abstraction versus mere memorization. To implement this measurement, we design a benchmark: systematic symbol remapping in rule-based tasks, which forces models to demonstrate genuine pattern recognition beyond superficial token matching. Extensive LLM evaluations using this benchmark (commercial API models, 7B-70B, multi-agent) reveal:1) critical limitations in non-decimal arithmetic and symbolic reasoning; 2) persistent abstraction gaps despite chain-of-thought prompting; and 3) \(\scoreDelta\)'s effectiveness in robustly measuring memory dependence by quantifying performance degradation under symbol remapping, particularly highlighting operand-specific memorization. These findings underscore that current LLMs, despite domain-specific strengths, still lack robust abstract reasoning, highlighting key areas for future improvement.
Ashery et al. recently argue that large language models (LLMs), when paired to play a classic "naming game," spontaneously develop linguistic conventions reminiscent of human social norms. Here, we show that their results are better explained by data leakage: the models simply reproduce conventions they already encountered during pre-training. Despite the authors' mitigation measures, we provide multiple analyses demonstrating that the LLMs recognize the structure of the coordination game and recall its outcomes, rather than exhibit "emergent" conventions. Consequently, the observed behaviors are indistinguishable from memorization of the training corpus. We conclude by pointing to potential alternative strategies and reflecting more generally on the place of LLMs for social science models.
We answer the question in the title, showing that adversarial training (AT) for diffusion models (DMs) fundamentally differs from classifiers: while AT in classifiers enforces output invariance, AT in DMs requires equivariance to keep the diffusion process aligned with the data distribution. AT is a way to enforce smoothness in the diffusion flow, improving robustness to outliers and corrupted data. Unlike prior art, our method makes no assumptions about the noise model and integrates seamlessly into diffusion training by adding random noise, similar to randomized smoothing, or adversarial noise, akin to AT. This enables intrinsic capabilities such as handling noisy data, dealing with extreme variability such as outliers, preventing memorization, and improving robustness. We rigorously evaluate our approach with proof-of-concept datasets with known distributions in low- and high-dimensional space, thereby taking a perfect measure of errors; we further evaluate on standard benchmarks such as CIFAR-10, CelebA and LSUN Bedroom, showing strong performance under severe noise, data corruption, and iterative adversarial attacks.
The rapid advancement of reasoning capabilities in large language models (LLMs) has led to notable improvements on mathematical benchmarks. However, many of the most commonly used evaluation datasets (e.g., AIME 2024) are widely available online, making it difficult to disentangle genuine reasoning from potential memorization. Furthermore, these benchmarks do not evaluate proof-writing capabilities, which are crucial for many mathematical tasks. To address this, we introduce MathArena, a new benchmark based on the following key insight: recurring math competitions provide a stream of high-quality, challenging problems that can be used for real-time evaluation of LLMs. By evaluating models as soon as new problems are released, we effectively eliminate the risk of contamination. Using this framework, we find strong signs of contamination in AIME 2024. Nonetheless, evaluations on harder competitions, such as SMT 2025 -- published well after model release dates -- demonstrate impressive reasoning capabilities in top-performing models. MathArena is also the first benchmark for proof-writing capabilities. On USAMO 2025, even top models score below 25%, far behind their performance on final-answer tasks. So far, we have evaluated 30 models across five competitions, totaling 149 problems. As an evolving benchmark, MathArena will continue to track the progress of LLMs on newly released competitions, ensuring rigorous and up-to-date evaluation of mathematical reasoning.
Large language models (LLMs) are known to memorize and recall English text from their pretraining data. However, the extent to which this ability generalizes to non-English languages or transfers across languages remains unclear. This paper investigates multilingual and cross-lingual memorization in LLMs, probing if memorized content in one language (e.g., English) can be recalled when presented in translation. To do so, we introduce OWL, a dataset of 31.5K aligned excerpts from 20 books in ten languages, including English originals, official translations (Vietnamese, Spanish, Turkish), and new translations in six low-resource languages (Sesotho, Yoruba, Maithili, Malagasy, Setswana, Tahitian). We evaluate memorization across model families and sizes through three tasks: (1) direct probing, which asks the model to identify a book's title and author; (2) name cloze, which requires predicting masked character names; and (3) prefix probing, which involves generating continuations. We find that LLMs consistently recall content across languages, even for texts without direct translation in pretraining data. GPT-4o, for example, identifies authors and titles 69% of the time and masked entities 6% of the time in newly translated excerpts. Perturbations (e.g., masking characters, shuffling words) modestly reduce direct probing accuracy (7% drop for shuffled official translations). Our results highlight the extent of cross-lingual memorization and provide insights on the differences between the models.
Recent work suggests that large language models (LLMs) can perform multi-hop reasoning implicitly -- producing correct answers without explicitly verbalizing intermediate steps -- but the underlying mechanisms remain poorly understood. In this paper, we study how such implicit reasoning emerges by training transformers from scratch in a controlled symbolic environment. Our analysis reveals a three-stage developmental trajectory: early memorization, followed by in-distribution generalization, and eventually cross-distribution generalization. We find that training with atomic triples is not necessary but accelerates learning, and that second-hop generalization relies on query-level exposure to specific compositional structures. To interpret these behaviors, we introduce two diagnostic tools: cross-query semantic patching, which identifies semantically reusable intermediate representations, and a cosine-based representational lens, which reveals that successful reasoning correlates with the cosine-base clustering in hidden space. This clustering phenomenon in turn provides a coherent explanation for the behavioral dynamics observed across training, linking representational structure to reasoning capability. These findings provide new insights into the interpretability of implicit multi-hop reasoning in LLMs, helping to clarify how complex reasoning processes unfold internally and offering pathways to enhance the transparency of such models.
Neural networks transform high-dimensional data into compact, structured representations, often modeled as elements of a lower dimensional latent space. In this paper, we present an alternative interpretation of neural models as dynamical systems acting on the latent manifold. Specifically, we show that autoencoder models implicitly define a latent vector field on the manifold, derived by iteratively applying the encoding-decoding map, without any additional training. We observe that standard training procedures introduce inductive biases that lead to the emergence of attractor points within this vector field. Drawing on this insight, we propose to leverage the vector field as a representation for the network, providing a novel tool to analyze the properties of the model and the data. This representation enables to: (i) analyze the generalization and memorization regimes of neural models, even throughout training; (ii) extract prior knowledge encoded in the network's parameters from the attractors, without requiring any input data; (iii) identify out-of-distribution samples from their trajectories in the vector field. We further validate our approach on vision foundation models, showcasing the applicability and effectiveness of our method in real-world scenarios.
Diffusion models now set the benchmark in high-fidelity generative sampling, yet they can, in principle, be prone to memorization. In this case, their learned score overfits the finite dataset so that the reverse-time SDE samples are mostly training points. In this paper, we interpret the empirical score as a noisy version of the true score and show that its covariance matrix is asymptotically a re-weighted data PCA. In large dimension, the small time limit makes the noise variance blow up while simultaneously reducing spatial correlation. To reduce this variance, we introduce a kernel-smoothed empirical score and analyze its bias-variance trade-off. We derive asymptotic bounds on the Kullback-Leibler divergence between the true distribution and the one generated by the modified reverse SDE. Regularization on the score has the same effect as increasing the size of the training dataset, and thus helps prevent memorization. A spectral decomposition of the forward diffusion suggests better variance control under some regularity conditions of the true data distribution. Reverse diffusion with kernel-smoothed empirical score can be reformulated as a gradient descent drifted toward a Log-Exponential Double-Kernel Density Estimator (LED-KDE). This perspective highlights two regularization mechanisms taking place in denoising diffusions: an initial Gaussian kernel first diffuses mass isotropically in the ambient space, while a second kernel applied in score space concentrates and spreads that mass along the data manifold. Hence, even a straightforward regularization-without any learning-already mitigates memorization and enhances generalization. Numerically, we illustrate our results with several experiments on synthetic and MNIST datasets.