Modern neural language models that are widely used in various NLP tasks risk memorizing sensitive information from their training data.Understanding this memorization is important in real world applications and also from a learning-theoretical perspective. An open question in previous studies of language model memorization is how to filter out common'' memorization. In fact, most memorization criteria strongly correlate with the number of occurrences in the training set, capturing memorized familiar phrases, public knowledge, templated texts, or other repeated data.We formulate a notion of counterfactual memorization which characterizes how a model's predictions change if a particular document is omitted during training.We identify and study counterfactually-memorized training examples in standard text datasets.We estimate the influence of each memorized training example on the validation set and on generated texts, showing how this can provide direct evidence of the source of memorization at test time.

Despite their wide adoption, the underlying training and memorization dynamics of very large language models is not well understood. We empirically study exact memorization in causal and masked language modeling, across model sizes and throughout the training process. We measure the effects of dataset size, learning rate, and model size on memorization, finding that larger language models memorize training data faster across all settings. Surprisingly, we show that larger models can memorize a larger portion of the data before over-fitting and tend to forget less throughout the training process. We also analyze the memorization dynamics of different parts of speech and find that models memorize nouns and numbers first; we hypothesize and provide empirical evidence that nouns and numbers act as a unique identifier for memorizing individual training examples.

Memorization, or the tendency of large language models (LLMs) to output entire sequences from their training data verbatim, is a key concern for deploying language models. In particular, it is vital to minimize a model's memorization of sensitive datapoints such as those containing personal identifiable information (PII). The prevalence of such undesirable memorization can pose issues for model trainers, and may even require discarding an otherwise functional model. We therefore seek to predict which sequences will be memorized before a large model's full train-time by extrapolating the memorization behavior of lower-compute trial runs. We measure memorization in the Pythia model suite and plot scaling laws for forecasting memorization, allowing us to provide equi-compute recommendations to maximize the reliability (recall) of such predictions.

Prompt learning approaches have made waves in natural language processing by inducing better few-shot performance while they still follow a parametric-based learning paradigm; the oblivion and rote memorization problems in learning may encounter unstable generalization issues. Specifically, vanilla prompt learning may struggle to utilize atypical instances by rote during fully-supervised training or overfit shallow patterns with low-shot data. To alleviate such limitations, we develop RetroPrompt with the motivation of decoupling knowledge from memorization to help the model strike a balance between generalization and memorization. In contrast with vanilla prompt learning, RetroPrompt constructs an open-book knowledge-store from training instances and implements a retrieval mechanism during the process of input, training and inference, thus equipping the model with the ability to retrieve related contexts from the training corpus as cues for enhancement. Extensive experiments demonstrate that RetroPrompt can obtain better performance in both few-shot and zero-shot settings.

Machine learning models often rely on simple spurious features -- patterns in training data that correlate with targets but are not causally related to them, like image backgrounds in foreground classification. This reliance typically leads to imbalanced test performance across minority and majority groups. In this work, we take a closer look at the fundamental cause of such imbalanced performance through the lens of memorization, which refers to the ability to predict accurately on \textit{atypical} examples (minority groups) in the training set but failing in achieving the same accuracy in the testing set. This paper systematically shows the ubiquitous existence of spurious features in a small set of neurons within the network, providing the first-ever evidence that memorization may contribute to imbalanced group performance. Through three experimental sources of converging empirical evidence, we find the property of a small subset of neurons or channels in memorizing minority group information. Inspired by these findings, we articulate the hypothesis: the imbalanced group performance is a byproduct of ``noisy'' spurious memorization confined to a small set of neurons. To further substantiate this hypothesis, we show that eliminating these unnecessary spurious memorization patterns via a novel framework during training can significantly affect the model performance on minority groups. Our experimental results across various architectures and benchmarks offer new insights on how neural networks encode core and spurious knowledge, laying the groundwork for future research in demystifying robustness to spurious correlation.

We propose a novel framework to perform classification via deep learning in the presence of noisy annotations. When trained on noisy labels, deep neural networks have been observed to first fit the training data with clean labels during an "early learning" phase, before eventually memorizing the examples with false labels. We prove that early learning and memorization are fundamental phenomena in high-dimensional classification tasks, even in simple linear models, and give a theoretical explanation in this setting. Motivated by these findings, we develop a new technique for noisy classification tasks, which exploits the progress of the early learning phase. In contrast with existing approaches, which use the model output during early learning to detect the examples with clean labels, and either ignore or attempt to correct the false labels, we take a different route and instead capitalize on early learning via regularization. There are two key elements to our approach.

Large Language Models (LLMs) are prevalent in modern applications but often memorize training data, leading to privacy breaches and copyright issues. Existing research has mainly focused on posthoc analyses, such as extracting memorized content or developing memorization metrics, without exploring the underlying architectural factors that contribute to memorization. In this work, we investigate memorization from an architectural lens by analyzing how attention modules at different layers impact its memorization and generalization performance. Using attribution techniques, we systematically intervene in the LLM architecture by bypassing attention modules at specific blocks while keeping other components like layer normalization and MLP transformations intact. We provide theorems analyzing our intervention mechanism from a mathematical view, bounding the difference in layer outputs with and without our attributions. Our theoretical and empirical analyses reveal that attention modules in deeper transformer blocks are primarily responsible for memorization, whereas earlier blocks are crucial for the models generalization and reasoning capabilities. We validate our findings through comprehensive experiments on different LLM families (Pythia and GPTNeo) and five benchmark datasets. Our insights offer a practical approach to mitigate memorization in LLMs while preserving their performance, contributing to safer and more ethical deployment in real world applications.

Recently, Large Language Models (LLMs) have shown impressive performance in character understanding tasks, such as analyzing the roles, personalities, and relationships of fictional characters. However, the extensive pre-training corpora used by LLMs raise concerns that they may rely on memorizing popular fictional works rather than genuinely understanding and reasoning about them. In this work, we argue that 'gist memory'-capturing essential meaning - should be the primary mechanism for character understanding tasks, as opposed to 'verbatim memory' - exact match of a string. We introduce a simple yet effective method to mitigate mechanized memorization in character understanding evaluations while preserving the essential implicit cues needed for comprehension and reasoning. Our approach reduces memorization-driven performance on popular fictional works from 96% accuracy to 72% and results in up to an 18% drop in accuracy across various character understanding tasks. These findings underscore the issue of data contamination in existing benchmarks, which often measure memorization rather than true character understanding.

In this paper, we explore the foundational mechanisms of memorization and generalization in Large Language Models (LLMs), inspired by the functional specialization observed in the human brain. Our investigation serves as a case study leveraging specially designed datasets and experimental-scale LLMs to lay the groundwork for understanding these behaviors. Specifically, we aim to first enable LLMs to exhibit both memorization and generalization by training with the designed dataset, then (a) examine whether LLMs exhibit neuron-level spatial differentiation for memorization and generalization, (b) predict these behaviors using model internal representations, and (c) steer the behaviors through inference-time interventions. Our findings reveal that neuron-wise differentiation of memorization and generalization is observable in LLMs, and targeted interventions can successfully direct their behavior.

Tabular data generation has attracted significant research interest in recent years, with the tabular diffusion models greatly improving the quality of synthetic data. However, while memorization, where models inadvertently replicate exact or near-identical training data, has been thoroughly investigated in image and text generation, its effects on tabular data remain largely unexplored. In this paper, we conduct the first comprehensive investigation of memorization phenomena in diffusion models for tabular data. Our empirical analysis reveals that memorization appears in tabular diffusion models and increases with larger training epochs. We further examine the influence of factors such as dataset sizes, feature dimensions, and different diffusion models on memorization. Additionally, we provide a theoretical explanation for why memorization occurs in tabular diffusion models. To address this issue, we propose TabCutMix, a simple yet effective data augmentation technique that exchanges randomly selected feature segments between random same-class training sample pairs. Building upon this, we introduce TabCutMixPlus, an enhanced method that clusters features based on feature correlations and ensures that features within the same cluster are exchanged together during augmentation. This clustering mechanism mitigates out-of-distribution (OOD) generation issues by maintaining feature coherence. Experimental results across various datasets and diffusion models demonstrate that TabCutMix effectively mitigates memorization while maintaining high-quality data generation.