The evolution of pre-trained large language models (LLMs) has significantly transformed natural language processing.However, these advancements pose challenges, particularly the unintended memorization of training data, which raises ethical and privacy concerns.While prior research has largely focused on mitigating memorization or extracting memorized information, the deliberate control of memorization has been underexplored.This study addresses this gap by introducing a novel and unified gradient-based weight pruning framework to freely control memorization rates in LLMs.Our method enables fine-grained control over pruning parameters, allowing models to suppress or enhance memorization based on application-specific requirements.Experimental results demonstrate that our approach effectively balances the trade-offs between memorization and generalization, with an increase of up to 89.3% in Fractional ER suppression and 40.9% in Exact ER amplification compared to the original models.

Group Ordered Weighted L1-Norm (Group OWL) regularized models have emerged as a useful procedure for high-dimensional sparse multi-task learning with correlated features. Proximal gradient methods are used as standard approaches to solving Group OWL models. However, Group OWL models usually suffer huge computational costs and memory usage when the feature size is large in the high-dimensional scenario. To address this challenge, in this paper, we are the first to propose the safe screening rule for Group OWL models by effectively tackling the structured non-separable penalty, which can quickly identify the inactive features that have zero coefficients across all the tasks. Thus, by removing the inactive features during the training process, we may achieve substantial computational gain and memory savings. More importantly, the proposed screening rule can be directly integrated with the existing solvers both in the batch and stochastic settings. Theoretically, we prove our screening rule is safe and also can be safely applied to the existing iterative optimization algorithms. Our experimental results demonstrate that our screening rule can effectively identify the inactive features and leads to a significant computational speedup without any loss of accuracy.

Shuyang Yu, Runxue Bao, Parminder Bhatia, Taha Kass-Hout, Jiayu Zhou, Cao Xiao. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

LLM-as-a-Judge has emerged as a promising tool for automatically evaluating generated outputs, but its reliability is often undermined by potential biases in judgment. Existing efforts to mitigate these biases face key limitations: in-context learning-based methods fail to address rooted biases due to the evaluator’s limited capacity for self-reflection, whereas fine-tuning is not applicable to all evaluator types, especially closed-source models. To address this challenge, we introduce the **R**easoning-based **B**ias **D**etector (RBD), which is a plug-in module that identifies biased evaluations and generates structured reasoning to guide evaluator self-correction. Rather than modifying the evaluator itself, RBD operates externally and engages in an iterative process of bias detection and feedback-driven revision. To support its development, we design a complete pipeline consisting of biased dataset construction, supervision collection, distilled reasoning-based fine-tuning of RBD, and integration with LLM evaluators. We fine-tune four sizes of RBD models, ranging from 1.5B to 14B, and observe consistent performance improvements across all scales. Experimental results on 4 bias types—verbosity, position, bandwagon, and sentiment—evaluated using 8 LLM evaluators demonstrate RBD’s strong effectiveness. For example, the RBD-8B model improves evaluation accuracy by an average of 18.5% and consistency by 10.9%, and surpasses prompting-based baselines and fine-tuned judges by 12.8% and 17.2%, respectively. These results highlight RBD’s effectiveness and scalability. Additional experiments further demonstrate its strong generalization across biases and domains, as well as its efficiency.

Environmental ecosystems exhibit complex and evolving dynamics over time, making the modeling of non-stationary processes critically important. However, traditional methods often rely on static models trained on entire datasets, failing to capture the non-stationary and drastically fluctuating characteristics. Dynamically adjusting models to evolving data is challenging, as they can easily either lag behind new trends or overfit newly received data. To address these challenges, we propose Domain-Adaptive Continual Meta-Learning (DACM) method, aiming to automatically detect distribution shifts and adapt to newly emergent domains. In particular, while DACM continuously explores the sequential temporal data, it also exploits historical data that are similar in distribution to the current observations. By striking a balance between temporal exploration and distributional exploitation, DACM quickly adjusts the model to stay up-to-date with new trends while maintaining generalization ability to data with similar distributions. We demonstrate the effectiveness of DACM on a real-world water temperature prediction dataset, where it outperforms diverse baseline models and shows strong adaptability and predictive performance in non-stationary environments.

Current techniques for deep neural network (DNN) pruning often involve intricate multi-step processes that re-quire domain-specific expertise, making their widespread adoption challenging. To address the limitation, the Only-Train-Once (OTO) and OTOv2 are proposed to eliminate the need for additional fine-tuning steps by directly training and compressing a general DNN from scratch. Never-theless, the static design of optimizers (in OTO) can lead to convergence issues of local optima. In this paper, we proposed the Auto-Train-Once (A TO), an innovative net-work pruning algorithm designed to automatically reduce the computational and storage costs of DNNs. During the model training phase, our approach not only trains the tar-get model but also leverages a controller network as an ar-chitecture generator to guide the learning of target model weights. Furthermore, we developed a novel stochastic gradient algorithm that enhances the coordination between model training and controller network training, thereby im-proving pruning performance. We provide a comprehen-sive convergence analysis as well as extensive experiments, and the results show that our approach achieves state-of-the-art performance across various model architectures (including ResNet18, ResNet34, ResNet50, ResNet56, and MobileNetv2) on standard benchmark datasets (CIFAR-10, CIFAR-100, and ImageNet). The code is available at https: 11 g i thub. comlxidon gwul Auto Train Once.

Zhepeng Wang, Runxue Bao, Yawen Wu, Jackson Taylor, Cao Xiao, Feng Zheng, Weiwen Jiang, Shangqian Gao, Yanfu Zhang. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Pairwise learning is an important machine-learning topic with many practical applications. An online algorithm is the first choice for processing streaming data and is preferred for handling large-scale pairwise learning problems. However, existing online pairwise learning algorithms are not scalable and efficient enough for large-scale high-dimensional data, because they were designed based on singly stochastic gradients. To address this challenging problem, in this article, we propose a dynamic doubly stochastic gradient algorithm (D2SG) for online pairwise learning. Especially, only the time and space complexities of are needed for incorporating a new sample, where is the dimensionality of data. This means that our D2SG is much faster and more scalable than the existing online pairwise learning algorithms while the statistical accuracy can be guaranteed through our rigorous theoretical analysis under standard assumptions. The experimental results on a variety of real-world datasets not only confirm the theoretical result of our new D2SG algorithm, but also show that D2SG has better efficiency and scalability than the existing online pairwise learning algorithms.

We provide both empirical and theoretical insights to demystify the gravity well phenomenon in the optimization landscape. We start from describe the problem setup and theoretical results (an escape time lower bound) of the Softmax Gravity Well (SGW) in the literature. Then we move toward the understanding of a recent observation called ASR gravity well. We provide an explanation of why normal distribution with high variance can lead to suboptimal plateaus from an energy function point of view. We also contribute to the empirical insights of curriculum learning by comparison of policy initialization by different normal distributions. Furthermore, we provide the ASR escape time lower bound to understand the ASR gravity well theoretically. Future work includes more specific modeling of the reward as a function of time and quantitative evaluation of normal distribution’s influence on policy initialization.

Sparsity regularized loss minimization problems play an important role in various fields including machine learning, data mining, and modern statistics. Proximal gradient descent method and coordinate descent method are the most popular approaches to solving the minimization problem. Although existing methods can achieve implicit model identification, aka support set identification, in a finite number of iterations, these methods still suffer from huge computational costs and memory burdens in high-dimensional scenarios. The reason is that the support set identification in these methods is implicit and thus cannot explicitly identify the low-complexity structure in practice, namely, they cannot discard useless coefficients of the associated features to achieve algorithmic acceleration via dimension reduction. To address this challenge, we propose a novel accelerated doubly stochastic gradient descent (ADSGD) method for sparsity regularized loss minimization problems, which can reduce the number of block iterations by eliminating inactive coefficients during the optimization process and eventually achieve faster explicit model identification and improve the algorithm efficiency. Theoretically, we first prove that ADSGD can achieve a linear convergence rate and lower overall computational complexity. More importantly, we prove that ADSGD can achieve a linear rate of explicit model identification. Numerically, experimental results on benchmark datasets confirm the efficiency of our proposed method.

There is some recent research interest in algorithmic fairness for biased data. There are a variety of pre-, in-, and post-processing methods designed for this problem. However, these methods are exclusively targeting data unfairness and algorithmic unfairness. In this paper, we propose a novel intra-processing method to broaden the application scenario of fairness methods, which can simultaneously address the two bias sources. Since training modern deep models from scratch is expensive due to the enormous training data and the complicated structures, we propose an augmentation and fine-tuning framework. First, we design an adversarial attack to generate weighted samples disentangled with the protected attribute. Next, we identify the fair sub-structure in the biased model and fine-tune the model via weight reactivation. At last, we provide an optional joint training scheme for the augmentation and the fine-tuning. Our method can be combined with a variety of fairness measures. We benchmark our method and some related baselines to show the advantage and the scalability. Experimental results on several standard datasets demonstrate that our approach can effectively learn fair augmentation and achieve superior results to the state-of-the-art baselines. Our method also generalizes well to different types of data.

Parallel optimization has become popular for large-scale learning in the past decades. However, existing methods suffer from huge computational costs, memory usage, and communication burden in high-dimensional scenarios. To address the challenges, we propose a new accelerated doubly sparse asynchronous learning (DSAL) method for stochastic composite optimization, under which two algorithms are proposed on shared-memory and distributed-memory architecture respectively, which only conducts gradient descent on the nonzero coordinates (data sparsity) and active set (model sparsity). The proposed algorithm can converge much faster and achieve significant speedup by simultaneously enjoying the sparsity of the model and data. Moreover, by sending the gradients on the active set only, communication costs are dramatically reduced. Theoretically, we prove that the proposed method achieves the linear convergence rate with lower overall complexity and can achieve the model identification in a finite number of iterations almost surely. Finally, extensive experimental results on benchmark datasets confirm the superiority of our proposed method.

Ordered Weighted $L_{1}$ (OWL) regularized regression is a new regression analysis for high-dimensional sparse learning. Proximal gradient methods are used as standard approaches to solve OWL regression. However, it is still a burning issue to solve OWL regression due to considerable computational cost and memory usage when the feature or sample size is large. In this paper, we propose the first safe screening rule for OWL regression by exploring the order of the primal solution with the unknown order structure via an iterative strategy, which overcomes the difficulties of tackling the non-separable regularizer. It effectively avoids the updates of the parameters whose coefficients must be zero during the learning process. More importantly, the proposed screening rule can be easily applied to standard and stochastic proximal gradient methods. Moreover, we prove that the algorithms with our screening rule are guaranteed to have identical results with the original algorithms. Experimental results on a variety of datasets show that our screening rule leads to a significant computational gain without any loss of accuracy, compared to existing competitive algorithms.

Variable selection is a challenging problem in high-dimensional linear regression problems with a large number of predictors. Thus, sparsity-inducing and clustering-inducing regularization methods are widely used to identify highly correlated covariates. Ordered Weight L1(OWL) family of regularizers for linear regression perform well to identify precise clusters of correlated covariates and interpret the effect of each variable. Solution path algorithms are helpful to select hyperparameters to tune the OWL model. Due to over-complex representation of the penalty, so far the OWL model has no solution path algorithms for hyperparameter selection. To address this challenge, in this paper, we propose an efficient approximate solution path algorithm (OWLAGPath) to solve the OWL model with accuracy guarantee. For a given accuracy bound ε, OWLAGPath can find the corresponding solutions for the OWL model with numerous hyperparameters while keeping the sparsity and precise features grouping properties. Theoretically, we prove that all the solutions produced by OWLAGPath can strictly satisfy the given accuracy bound ε. The experimental results on three benchmark datasets not only confirm the effectiveness and efficiency of our OWLAGPath algorithm, but also show the advantages of OWLAGPath for model selection than the existing algorithms.