Ziyan Wang, Enmao Diao, Qi Le, Pu Wang, Minwoo Lee, Shu-ping Yeh, Evgeny Stupachenko, Hao Feng, Li Yang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

We introduce Probe Pruning (PP), a novel framework for online, dynamic, structured pruning of Large Language Models (LLMs) applied in a batch-wise manner. PP leverages the insight that not all samples and tokens contribute equally to the model's output, and probing a small portion of each batch effectively identifies crucial weights, enabling tailored dynamic pruning for different batches. It comprises three main stages: probing, history-informed pruning, and full inference. In the probing stage, PP selects a small yet crucial set of hidden states, based on residual importance, to run a few model layers ahead. During the history-informed pruning stage, PP strategically integrates the probing states with historical states. Subsequently, it structurally prunes weights based on the integrated states and the PP importance score, a metric developed specifically to assess the importance of each weight channel in maintaining performance. In the final stage, full inference is conducted on the remaining weights. A major advantage of PP is its compatibility with existing models, as it operates without requiring additional neural network modules or fine-tuning. Comprehensive evaluations of PP on LLaMA-2/3 and OPT models reveal that even minimal probing—using just 1.5% of FLOPs—can substantially enhance the efficiency of structured pruning of LLMs. For instance, when evaluated on LLaMA-2-7B with WikiText2, PP achieves a 2.56 times lower ratio of performance degradation per unit of latency reduction compared to the state-of-the-art method at a 40\% pruning ratio.

Ensuring that generative AI systems align with human values is essential but challenging, especially when considering multiple human values and their potential trade-offs. Since human values can be personalized and dynamically change over time, the desirable levels of value alignment vary across different ethnic groups, industry sectors, and user cohorts. Within existing frameworks, it is hard to define human values and align AI systems accordingly across different directions simultaneously, such as harmlessness, helpfulness, and positiveness. To address this, we develop a novel, first-principle approach called Multi-Human-Value Alignment Palette (MAP), which navigates the alignment across multiple human values in a structured and reliable way. MAP formulates the alignment problem as an optimization task with user-defined constraints, which define human value targets. It can be efficiently solved via a primal-dual approach, which determines whether a user-defined alignment target is achievable and how to achieve it. We conduct a detailed theoretical analysis of MAP by quantifying the trade-offs between values, the sensitivity to constraints, the fundamental connection between multi-value alignment and sequential alignment, and proving that linear weighted rewards are sufficient for multi-value alignment. Extensive experiments demonstrate MAP's ability to align multiple values in a principled manner while delivering strong empirical performance across various tasks.

Xinran Wang, Enmao Diao, Qi Le, Jie Ding, Ali Anwar. 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.

Large language models can generate rare but catastrophic outputs, such as harmful conversations or insecure code. Existing Reinforcement Learning from Human Feedback (RLHF) typically maximizes average reward, leaving high-risk tail events insufficiently controlled. We introduce Quantile‑Guided Alignment (QA), a framework that allows users to specify desired improvements at any quantile—individually or across multiple reward dimensions—thus shifting the distribution of outputs with finer control toward safer, more desirable outcomes. The method extends standard RLHF via an augmented reward formulation that enforces quantile constraints. Experiments on conversation and code‐generation tasks show that quantile alignment significantly enhances quality at targeted tails while maintaining overall performance. The results position QA as a principled route to risk‑calibrated language models with tail‑focused alignment.

Methods in the field of quickest change detection rapidly detect in real-time a change in the data-generating distribution of an online data stream. Existing methods have been able to detect this change point when the densities of the pre-and post-change distributions are known. Recent work has extended these results to the case where the pre-and post-change distributions are known only by their score functions. This work considers the case where the pre-and post-change score functions are known only to correspond to distributions in two disjoint sets. This work selects a pair of least-favorable distributions from these sets to robustify the existing score-based quickest change detection algorithm, the properties of which are studied. This paper calculates the least-favorable distributions for specific model classes and provides methods of estimating the leastfavorable distributions for common constructions. Simulation results are provided demonstrating the performance of our robust change detection algorithm.

Federated Learning (FL) is a collaborative machine learning framework that allows multiple users to train models utilizing their local data in a distributed manner. However, considerable statistical heterogeneity in local data across devices often leads to suboptimal model performance compared with independently and identically distributed (IID) data scenarios. In this paper, we introduce DynamicFL, a new FL framework that investigates the trade-offs between global model performance and communication costs for two widely adopted FL methods: Federated Stochastic Gradient Descent (FedSGD) and Federated Averaging (FedAvg). Our approach allocates diverse communication resources to clients based on their data statistical heterogeneity, considering communication resource constraints, and attains substantial performance enhancements compared to uniform communication resource allocation. Notably, our method bridges the gap between FedSGD and FedAvg, providing a flexible framework leveraging communication heterogeneity to address statistical heterogeneity in FL. Through extensive experiments, we demonstrate that DynamicFL surpasses current state-of-the-art methods with up to a 10% increase in model accuracy, demonstrating its adaptability and effectiveness in tackling data statistical heterogeneity challenges.

Neural speech codecs aim to compress input signals into minimal bits while maintaining content quality in a low-latency manner.However, existing neural codecs often trade model complexity for reconstruction performance.These codecs primarily use convolutional blocks for feature transformation, which are not inherently suited for capturing the local redundancies in speech signals.To compensate, they require either adversarial discriminators or a large number of model parameters to enhance audio quality.In response to these challenges, we introduce the Efficient Speech Codec (ESC) 1 , a lightweight, parameter-efficient speech codec based on a cross-scale residual vector quantization scheme and transformers.Our model employs mirrored hierarchical window transformer blocks and performs step-wise decoding from coarse-tofine feature representations.To enhance bitrate efficiency, we propose a novel combination of vector quantization techniques along with a pre-training paradigm.Extensive experiments demonstrate that ESC can achieve high-fidelity speech reconstruction with significantly lower model complexity, making it a promising alternative to existing convolutional audio codecs.

In GPR B-scan images, ground surface clutter is the main source of interference, often obscuring or distorting subsurface target signals. We propose a deep autoencoder-based method to mitigate rough surface clutter in GPR images by treating it as an anomaly detection problem. Firstly, the rough surface region in a B-scan image is partitioned into small patches, which act as the training dataset for the deep autoencoder. Through the training process, the autoencoder learns and captures the patterns associated with the rough surface patches. Following training, the entire B-scan image is divided into small patches of the same size as the training patches, and each of them is fed into the autoencoder to compute an anomaly score. To reconstruct a clutter-reduced B-scan image, we employ a weighted sum approach to aggregate all patches based on their anomaly scores. We evaluate our method against conventional subspace projection techniques using simulated and field-collected B-scans. The results clearly indicate that our approach surpasses these subspace methods. Furthermore, we employ t-SNE analysis to gain deeper insights into our method’s effectiveness in reducing rough surface clutter. The outcomes of this analysis reinforce the practical viability of our approach for GPR image processing.

Classical quickest change detection algorithms require modeling pre-change and post-change distributions. Such an approach may not be feasible for various machine learning models because of the complexity of computing the explicit distributions. Additionally, these methods may suffer from a lack of robustness to model mismatch and noise. This paper develops a new variant of the classical Cumulative Sum (CUSUM) algorithm for the quickest change detection. This variant is based on Fisher divergence and the Hyvärinen score and is called the Hyvärinen score-based CUSUM (SCUSUM) algorithm. The SCUSUM algorithm allows the applications of change detection for unnormalized statistical models, i.e., models for which the probability density function contains an unknown normalization constant. The asymptotic optimality of the proposed algorithm is investigated by deriving expressions for average detection delay and the mean running time to a false alarm. Numerical results are provided to demonstrate the performance of the proposed algorithm.

Classical change detection algorithms typically require modeling pre-change and post-change distributions. The calculations may not be feasible for various machine learning models because of the complexity of computing the partition functions and normalized distributions. Additionally, these methods may suffer from a lack of robustness to model mismatch and noise. In this paper, we develop a new variant of the classical Cumulative Sum (CUSUM) change detection, namely Score-based CUSUM (SCUSUM), based on Fisher divergence and the Hyvärinen score. Our method allows the applications of the quickest change detection for unnormalized distributions. We provide a theoretical analysis of the detection delay given the constraints on false alarms. We prove the asymptotic optimality of the proposed method in some particular cases. We also provide numerical experiments to demonstrate our method’s computation, performance, and robustness advantages.

Detecting an abrupt and persistent change in the underlying distribution of online data streams is an important problem in many applications. This paper proposes a new robust score-based algorithm called RSCUSUM, which can be applied to unnormalized models and addresses the issue of unknown post-change distributions. RSCUSUM replaces the Kullback-Leibler divergence with the Fisher divergence between pre- and post-change distributions for computational efficiency in unnormalized statistical models and introduces a notion of the “least favorable” distribution for robust change detection. The algorithm and its theoretical analysis are demonstrated through simulation studies.

Analyzing large-scale data from simulations of turbulent flows is memory intensive, requiring significant resources. This major challenge highlights the need for data compression techniques. In this study, we apply a physics-informed Deep Learning technique based on vector quantization to generate a discrete, low-dimensional representation of data from simulations of three-dimensional turbulent flows. The deep learning framework is composed of convolutional layers and incorporates physical constraints on the flow, such as preserving incompressibility and global statistical characteristics of the velocity gradients. The accuracy of the model is assessed using statistical, comparison-based similarity and physics-based metrics. The training data set is produced from Direct Numerical Simulation of an incompressible, statistically stationary, isotropic turbulent flow. The performance of this lossy data compression scheme is evaluated not only with unseen data from the stationary, isotropic turbulent flow, but also with data from decaying isotropic turbulence, and a Taylor-Green vortex flow. Defining the compression ratio (CR) as the ratio of original data size to the compressed one, the results show that our model based on vector quantization can offer CR = 85 with a mean square error (MSE) of$O$(10−3), and predictions that faithfully reproduce the statistics of the flow, except at the very smallest scales where there is some loss. Compared to the recent study based on a conventional autoencoder where compression is performed in a continuous space, our model improves the CR by more than 30 percent, and reduces the MSE by an order of magnitude. Our compression model is an attractive solution for situations where fast, high quality and low-overhead encoding and decoding of large data are required.

Collaborations among multiple organizations, such as financial institutions, medical centers, and retail markets in decentralized settings are crucial to providing improved service and performance. However, the underlying organizations may have little interest in sharing their local data, models, and objective functions. These requirements have created new challenges for multi-organization collaboration. In this work, we propose Gradient Assisted Learning (GAL), a new method for multiple organizations to assist each other in supervised learning tasks without sharing local data, models, and objective functions. In this framework, all participants collaboratively optimize the aggregate of local loss functions, and each participant autonomously builds its own model by iteratively fitting the gradients of the overarching objective function. We also provide asymptotic convergence analysis and practical case studies of GAL. Experimental studies demonstrate that GAL can achieve performance close to centralized learning when all data, models, and objective functions are fully disclosed.

Federated Learning allows the training of machine learning models by using the computation and private data resources of many distributed clients. Most existing results on Federated Learning (FL) assume the clients have ground-truth labels. However, in many practical scenarios, clients may be unable to label task-specific data due to a lack of expertise or resource. We propose SemiFL to address the problem of combining communication-efficient FL such as FedAvg with Semi-Supervised Learning (SSL). In SemiFL, clients have completely unlabeled data and can train multiple local epochs to reduce communication costs, while the server has a small amount of labeled data. We provide a theoretical understanding of the success of data augmentation-based SSL methods to illustrate the bottleneck of a vanilla combination of communication-efficient FL with SSL. To address this issue, we propose alternate training to 'fine-tune global model with labeled data' and 'generate pseudo-labels with the global model.' We conduct extensive experiments and demonstrate that our approach significantly improves the performance of a labeled server with unlabeled clients training with multiple local epochs. Moreover, our method outperforms many existing SSFL baselines and performs competitively with the state-of-the-art FL and SSL results.

A central issue of many statistical learning problems is to select an appropriate model from a set of candidate models. Large models tend to inflate the variance (or overfitting), while small models tend to cause biases (or underfitting) for a given fixed dataset. In this work, we address the critical challenge of model selection to strike a balance between model fitting and model complexity, thus gaining reliable predictive power. We consider the task of approaching the theoretical limit of statistical learning, meaning that the selected model has the predictive performance that is as good as the best possible model given a class of potentially misspecified candidate models. We propose a generalized notion of Takeuchi's information criterion and prove that the proposed method can asymptotically achieve the optimal out-sample prediction loss under reasonable assumptions. It is the first proof of the asymptotic property of Takeuchi's information criterion to our best knowledge. Our proof applies to a wide variety of nonlinear models, loss functions, and high dimensionality (in the sense that the models' complexity can grow with sample size). The proposed method can be used as a computationally efficient surrogate for leave-one-out cross-validation. Moreover, for modeling streaming data, we propose an online algorithm that sequentially expands the model complexity to enhance selection stability and reduce computation cost. Experimental studies show that the proposed method has desirable predictive power and significantly less computational cost than some popular methods.

We propose a new architecture for distributed image compression from a group of distributed data sources. The work is motivated by practical needs of data-driven codec design, low power consumption, robustness, and data privacy. The proposed architecture, which we refer to as Distributed Recurrent Autoencoder for Scalable Image Compression (DRASIC), is able to train distributed encoders and one joint decoder on correlated data sources. Its compression capability is much better than the method of training codecs separately. Meanwhile, the performance of our distributed system with 10 distributed sources is only within 2 dB peak signal-to-noise ratio (PSNR) of the performance of a single codec trained with all data sources. We experiment distributed sources with different correlations and show how our data-driven methodology well matches the Slepian-Wolf Theorem in Distributed Source Coding (DSC). To the best of our knowledge, this is the first data-driven DSC framework for general distributed code design with deep learning.

Motivated by the ever-increasing demands for limited communication bandwidth and low-power consumption, we propose a new methodology, named joint Variational Autoencoders with Bernoulli mixture models (VAB), for performing clustering in the compressed data domain. The idea is to reduce the data dimension by Variational Autoencoders (VAEs) and group data representations by Bernoulli mixture models (BMMs). Once jointly trained for compression and clustering, the model can be decomposed into two parts: a data vendor that encodes the raw data into compressed data, and a data consumer that classifies the received (compressed) data. In this way, the data vendor benefits from data security and communication bandwidth, while the data consumer benefits from low computational complexity. To enable training using the gradient descent algorithm, we propose to use the Gumbel-Softmax distribution to resolve the infeasibility of the back-propagation algorithm when assessing categorical samples.

Speech Emotion Recognition (SER) has emerged as a critical component of the next generation of human-machine interfacing technologies. In this work, we propose a new dual-level model that predicts emotions based on both MFCC features and mel-spectrograms produced from raw audio signals. Each utterance is preprocessed into MFCC features and two mel-spectrograms at different time-frequency resolutions. A standard LSTM processes the MFCC features, while a novel LSTM architecture, denoted as Dual-Sequence LSTM (DS-LSTM), processes the two mel-spectrograms simultaneously. The outputs are later averaged to produce a final classification of the utterance. Our proposed model achieves, on average, a weighted accuracy of 72.7% and an unweighted accuracy of 73.3% - a 6% improvement over current state-of-the-art unimodal models - and is comparable with multimodal models that leverage textual information as well as audio signals.

Recurrent Neural Network (RNN) and its variations such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), have become standard building blocks for learning online data of sequential nature in many research areas, including natural language processing and speech data analysis. In this paper, we present a new methodology to significantly reduce the number of parameters in RNNs while maintaining performance that is comparable or even better than classical RNNs. The new proposal, referred to as Restricted Recurrent Neural Network (RRNN), restricts the weight matrices corresponding to the input data and hidden states at each time step to share a large proportion of parameters. The new architecture can be regarded as a compression of its classical counterpart, but it does not require pre-training or sophisticated parameter fine-tuning, both of which are major issues in most existing compression techniques. Experiments on natural language modeling show that compared with its classical counterpart, the restricted recurrent architecture generally produces comparable results at about 50% compression rate. In particular, the Restricted LSTM can outperform classical RNN with even less number of parameters.

Machine learning systems learn from and make predictions by building models from observed data. Because large models tend to overfit while small models tend to underfit for a given fixed dataset, a critical challenge is to select an appropriate model (e.g. set of variables/features). Model selection aims to strike a balance between the goodness of fit and model complexity, and thus to gain reliable predictive power. In this paper, we study a penalized model selection technique that asymptotically achieves the optimal expected prediction loss (referred to as the limit of learning) offered by a set of candidate models. We prove that the proposed procedure is both statistically efficient in the sense that it asymptotically approaches the limit of learning, and computationally efficient in the sense that it can be much faster than cross validation methods. Our theory applies for a wide variety of model classes, loss functions, and high dimensions (in the sense that the models' complexity can grow with data size). We released a python package with our proposed method for general usage like logistic regression and neural networks.