论文检索

Deep Learning · Large Language Models

Haoyun Jiang, Haolin li, jianwei zhang, Fei Huang, Qiang Hu, Minmin Sun, Shuai Xiao, Yong Li, Junyang Lin, Jiangchao Yao

Large language models (LLMs) have demonstrated strong capabilities in handling long-context tasks, but processing such long contexts remains challenging due to the substantial memory requirements and inference latency. In this work, we discover that certain attention heads exhibit sequential consistency in their attention patterns, which can be persistently identified using a coefficient-of-variation-based algorithm. Inspired by this observation, we propose CateKV, a hybrid KV cache method that retains only critical token information for consistent heads, thereby reducing KV cache size and computational overhead, while preserving the majority of KV pairs in adaptive heads to ensure high accuracy. We show the unique characteristics of our algorithm and its extension with existing acceleration methods. Comprehensive evaluations on long-context benchmarks show that, while maintaining accuracy comparable to full attention, CateKV reduces memory usage by up to $2.72\times$ and accelerates decoding by $2.18\times$ in single-sample inputs, and boosts throughput by $3.96\times$ in batch scenarios.

Deep Learning · Attention Mechanisms

Haocheng Xi, Shuo Yang, Yilong Zhao, Chenfeng Xu, Muyang Li, Xiuyu Li, Yujun Lin, Han Cai, Jintao Zhang, Dacheng Li 等

Diffusion Transformers (DiTs) dominate video generation but their high computational cost severely limits real-world applicability, usually requiring tens of minutes to generate a few seconds of video even on high-performance GPUs. This inefficiency primarily arises from the quadratic computational complexity of 3D full attention with respect to the context length. In this paper, we propose a training-free framework termed Sparse VideoGen (SVG) that leverages the inherent sparsity in 3D full attention to boost inference efficiency. We reveal that the attention heads can be dynamically classified into two groups depending on distinct sparse patterns: (1) Spatial Head, where only spatially-related tokens within each frame dominate the attention output, and (2) Temporal Head, where only temporally-related tokens across different frames dominate. Based on this insight, SVG proposes an online profiling strategy to capture the dynamic sparse patterns and predicts the type of attention head. Combined with a novel hardware-efficient tensor layout transformation and customized kernel implementations, SVG achieves up to 2.28$\times$ and 2.33$\times$ end-to-end speedup on CogVideoX-v1.5 and HunyuanVideo, respectively, while preserving generation quality. Our code will be open-sourced upon publication.

Deep Learning

Hyeonah Kim, Sanghyeok Choi, Jiwoo Son, Jinkyoo Park, Changhyun Kwon

Effective search methods are crucial for improving the performance of deep generative models at test time. In this paper, we introduce a novel test-time search method, Neural Genetic Search (NGS), which incorporates the evolutionary mechanism of genetic algorithms into the generation procedure of deep models. The core idea behind NGS is its crossover, which is defined as parent-conditioned generation using trained generative models. This approach offers a versatile and easy-to-implement search algorithm for deep generative models. We demonstrate the effectiveness and flexibility of NGS through experiments across three distinct domains: routing problems, adversarial prompt generation for language models, and molecular design.

Optimization

Michael Sander, Vincent Roulet, Tianlin Liu, Mathieu Blondel

Energy-based models (EBMs) offer a flexible framework for parameterizing probability distributions using neural networks.However, learning EBMs by exact maximum likelihood estimation (MLE) is generally intractable, due to the need to compute the partition function.In this paper, we propose a novel min-min formulation for approximately learning probabilistic EBMs in combinatorially-large discrete spaces, such as sets or permutations. Our key idea is to jointly learn both an energy model and its log-partition, parameterized as a neural network. Our approach not only provides a novel tractable objective criterion to learn EBMs by stochastic gradient descent (without relying on MCMC), but also a novel means to estimate the log-partition function on unseen data points.On the theoretical side, we show that our approach recovers the optimal MLE solution when optimizing in the space of continuous functions.Furthermore, we show that our approach naturally extends to the broader family of Fenchel-Young losses, allowing us to obtainthe first tractable method for optimizing the sparsemax loss in combinatorially-large spaces.We demonstrate our approach on multilabel classification and label ranking.

Applications · Computer Vision

Dong Li, Yidi Liu, Xueyang Fu, Jie Huang, Senyan Xu, Qi Zhu, Zheng-Jun Zha

Image deraining aims to remove rain streaks from rainy images and restore clear backgrounds. Currently, some research that employs the Fourier transform has proved to be effective for image deraining, due to it acting as an effective frequency prior for capturing rain streaks. However, despite there exists dependency of low frequency and high frequency in images, these Fourier-based methods rarely exploit the correlation of different frequencies for conjuncting their learning procedures, limiting the full utilization of frequency information for image deraining. Alternatively, the recently emerged Mamba technique depicts its effectiveness and efficiency for modeling correlation in various domains (e.g., spatial, temporal), and we argue that introducing Mamba into its unexplored Fourier spaces to correlate different frequencies would help improve image deraining. This motivates us to propose a new framework termed FourierMamba, which performs image deraining with Mamba in the Fourier space. Owing to the unique arrangement of frequency orders in Fourier space, the core of FourierMamba lies in the scanning encoding of different frequencies, where the low-high frequency order formats exhibit differently in the spatial dimension (unarranged in axis) and channel dimension (arranged in axis). Therefore, we design FourierMamba that correlates Fourier space information in the spatial and channel dimensions with distinct designs. Specifically, in the spatial dimension Fourier space, we introduce the zigzag coding to scan the frequencies to rearrange the orders from low to high frequencies, thereby orderly correlating the connections between frequencies; in the channel dimension Fourier space with arranged orders of frequencies in axis, we can directly use Mamba to perform frequency correlation and improve the channel information representation. Extensive experiments reveal that our method outperforms state-of-the-art methods both qualitatively and quantitatively.

Deep Learning · Large Language Models

Tianjian Li, Daniel Khashabi

Aligning language models with human preferences relies on pairwise preference datasets. While some studies suggest that on-policy data consistently outperforms off-policy data for preference learning, others indicate that the advantages of on-policy data are task-dependent, highlighting the need for a systematic exploration of their interplay.In this work, we show that on-policy and off-policy data offer complementary strengths: on-policy data is particularly effective for reasoning tasks like math and coding, while off-policy data performs better on subjective tasks such as creative writing and making personal recommendations. Guided by these findings, we introduce SimpleMix, an approach to combine the complementary strengths of on-policy and off-policy preference learning by simply mixing these two data sources. Our empirical results across diverse tasks and benchmarks demonstrate that SimpleMix substantially improves language model alignment. Specifically, SimpleMix improves upon on-policy DPO and off-policy DPO by an average of 6.03 on Alpaca Eval 2.0. Moreover, it surpasses prior approaches that are much more complex in combining on- and off-policy data, such as HyPO and DPO-Mix-P, by an average of 3.05. These findings validate the effectiveness and efficiency of SimpleMix for enhancing preference-based alignment.

General Machine Learning · Evaluation

Wenbin Wang, Yu Shi, Ziping Zhao

The emergence of multi-dimensional data presents significant challenges for traditional regression models based on matrices or vectors, particularly in capturing multi-directional correlations. In response, tensor regression has been proposed as a powerful framework for modeling linear relationships among multi-dimensional variables. In this paper, we introduce a high-dimensional tensor-response tensor regression model under low-dimensional structural assumptions, such as sparsity and low-rankness. Assuming the underlying tensor lies within an unknown low-dimensional subspace, we consider a least squares estimation framework with non-convex penalties. Theoretically, we derive general risk bounds for the resulting estimators and demonstrate that they achieve the oracle statistical rates under mild technical conditions. To compute the proposed estimators efficiently, we introduce an accelerated proximal gradient algorithm demonstrating rapid convergence in practice. Extensive experiments on synthetic and real-world datasets validate the effectiveness of the proposed regression model and showcase the practical utility of the theoretical findings.

Deep Learning · Foundation Models

Jeonghoon Kim, Byeongchan Lee, Cheonbok Park, Yeontaek Oh, Beomjun Kim, Taehwan Yoo, Seongjin Shin, Dongyoon Han, Jinwoo Shin, Kang Min Yoo

Selecting a layer normalization (LN) strategy that stabilizes training and speeds convergence in Transformers remains difficult, even for today’s large language models (LLM). We present a comprehensive analytical foundation for understanding how different LN strategies influence training dynamics in large-scale Transformers. Until recently, Pre-LN and Post-LN have long dominated practices despite their limitations in large-scale training. However, several open-source models have recently begun silently adopting a third strategy without much explanation. This strategy places normalization layer **peripherally** around sublayers, a design we term **Peri-LN**. While Peri-LN has demonstrated promising performance, its precise mechanisms and benefits remain almost unexplored. Our in-depth analysis delineates the distinct behaviors of LN strategies, showing how each placement shapes activation variance and gradient propagation. To validate our theoretical insight, we conduct extensive experiments on Transformers up to $3.2$B parameters, showing that Peri-LN consistently achieves more balanced variance growth, steadier gradient flow, and convergence stability. Our results suggest that Peri-LN warrants broader consideration for large-scale Transformer architectures, providing renewed insights into the optimal placement of LN.

Probabilistic Methods · Monte Carlo and Sampling Methods

Bin Yang, Xiaojie Wang

Generating samples from a high dimensional probability distribution is a fundamental task with wide-ranging applications in the area of scientific computing, statistics and machine learning. This article revisits the popular Langevin Monte Carlo (LMC) sampling algorithms and provides a non-asymptotic error analysis in $\mathcal{W}_2$-distance in a non-convex setting. In particular, we prove an error bound $O(\sqrt{d} h)$, which guarantees a mixing time $ \tilde{O} (\sqrt{d} \epsilon^{-1})$ to achieve the accuracy tolerance $\epsilon$, under certain log-smooth conditions and the assumption that the target distribution satisfies a log-Sobolev inequality, as opposed to the strongly log-concave condition used in (Li et al., 2019; 2022). This bound matches the best one in the strongly log-concave case and improves upon the best-known convergence rates in non-convex settings. To prove it, we establish a new framework of uniform-in-time convergence for discretizations of SDEs. Distinct from (Li et al., 2019; 2022), we start from the finite-time mean-square fundamental convergence theorem, which combined with uniform-in-time moment bounds of LMC and the exponential ergodicity of SDEs in the non-convex setting gives the desired uniform-in-time convergence. Our framework also applies to the case when the gradient of the potential $U$ is non-globally Lipschitz with superlinear growth, for which modified LMC samplers are proposed and analyzed, with a non-asymptotic error bound in $\mathcal{W}_2$-distance obtained. Numerical experiments corroborate the theoretical analysis.

Social Aspects · Privacy

Guy Kornowski, Daogao Liu, Kunal Talwar

We study differentially private (DP) optimization algorithms for stochastic and empirical objectives which are neither smooth nor convex, and propose methods that return a Goldstein-stationary point with sample complexity bounds that improve on existing works.We start by providing a single-pass $(\epsilon,\delta)$-DP algorithm that returns an $(\alpha,\beta)$-stationary point as long as the dataset is of size $\widetilde{\Omega}(\sqrt{d}/\alpha\beta^{3}+d/\epsilon\alpha\beta^{2})$, which is $\Omega(\sqrt{d})$ times smaller than the algorithm of Zhang et al. (2024) for this task, where $d$ is the dimension.We then provide a multi-pass polynomial time algorithm which further improves the sample complexity to $\widetilde{\Omega}\left(d/\beta^2+d^{3/4}/\epsilon\alpha^{1/2}\beta^{3/2}\right)$, by designing a sample efficient ERM algorithm, and proving that Goldstein-stationary points generalize from the empirical loss to the population loss.

Deep Learning · Sequential Models, Time series

Matthew Faw, Rajat Sen, Yichen Zhou, Abhimanyu Das

Motivated by the recent success of time-series foundation models for zero-shot forecasting, we present a methodology for _in-context fine-tuning_ of a time-series foundation model. In particular, we design a pretrained foundation model that can be prompted (at inference time) with multiple time-series examples, in order to forecast a target time-series into the future. Our foundation model is specifically trained to utilize examples from multiple related time-series in its context window (in addition to the history of the target time-series) to help it adapt to the specific distribution of the target domain at inference time. We show that such a foundation model that uses in-context examples at inference time can obtain much better performance on popular forecasting benchmarks compared to supervised deep learning methods, statistical models, and other time series foundation models. Interestingly, our in-context fine-tuning approach even matches the performance of a foundation model that is explicitly fine-tuned on the target domain.

Theory · Game Theory

Yudong Hu, Haoran Li, Congying Han, Tiande Guo, Bonan Li, Mingqiang Li

Solving the Nash equilibrium in normal-form games with large-scale strategy spaces presents significant challenges. Open-ended learning frameworks, such as PSRO and its variants, have emerged as effective solutions. However, these methods often lack an efficient metric for evaluating strategy improvement, which limits their effectiveness in approximating equilibria.In this paper, we introduce a novel evaluative metric called Advantage, which possesses desirable properties inherently connected to the Nash equilibrium, ensuring that each strategy update approaches equilibrium. Building upon this, we propose the Advantage Policy Space Response Oracle (A-PSRO), an innovative unified open-ended learning framework applicable to both zero-sum and general-sum games. A-PSRO leverages the Advantage as a refined evaluation metric, leading to a consistent learning objective for agents in normal-form games. Experiments showcase that A-PSRO significantly reduces exploitability in zero-sum games and improves rewards in general-sum games, outperforming existing algorithms and validating its practical effectiveness.

Optimization · Non-Convex

Jiaqi Leng, Bin Shi

With rapid advancements in machine learning, first-order algorithms have emerged as the backbone of modern optimization techniques, owing to their computational efficiency and low memory requirements. Recently, the connection between accelerated gradient methods and damped heavy-ball motion, particularly within the framework of Hamiltonian dynamics, has inspired the development of innovative quantum algorithms for continuous optimization. One such algorithm, Quantum Hamiltonian Descent (QHD), leverages quantum tunneling to escape saddle points and local minima, facilitating the discovery of global solutions in complex optimization landscapes. However, QHD faces several challenges, including slower convergence rates compared to classical gradient methods and limited robustness in highly non-convex problems due to the non-local nature of quantum states. Furthermore, the original QHD formulation primarily relies on function value information, which limits its effectiveness. Inspired by insights from high-resolution differential equations that have elucidated the acceleration mechanisms in classical methods, we propose an enhancement to QHD by incorporating gradient information, leading to what we call gradient-based QHD. This gradient-based QHD achieves faster convergence and significantly increases the likelihood of identifying global solutions. Numerical simulations on challenging problem instances demonstrate that this gradient-based QHD outperforms existing quantum and classical methods by at least an order of magnitude.

Deep Learning · Theory

Ningyuan Huang, Miguel Sarabia, Abhinav Moudgil, Pau Rodriguez, Luca Zappella, Federico Danieli

State-Space Models (SSMs), and particularly Mamba, have recently emerged as a promising alternative to Transformers. Mamba introduces input selectivity to its SSM layer (S6) and incorporates convolution and gating into its block definition. While these modifications do improve Mamba's performance over its SSM predecessors, it remains largely unclear how Mamba leverages the additional functionalities provided by input selectivity, and how these interact with the other operations in the Mamba architecture. In this work, we demystify the role of input selectivity in Mamba, investigating its impact on function approximation power, long-term memorization, and associative recall capabilities.In particular: (i) we prove that the S6 layer of Mamba can represent projections onto *Haar wavelets*, providing an edge over its Diagonal SSM (S4D) predecessor in approximating discontinuous functions commonly arising in practice; (ii) we show how the S6 layer can dynamically counteract memory decay; (iii) we provide analytical solutions to the MQAR associative recall task using the Mamba architecture with different mixers --- Mamba, Mamba-2, and S4D. We demonstrate the tightness of our theoretical constructions with empirical results on concrete tasks. Our findings offer a mechanistic understanding of Mamba and reveal opportunities for improvement.

General Machine Learning · Sequential, Network, and Time Series Modeling

Andrea Testa, Søren Hauberg, Tamim Asfour, Leonel Rozo

Accurately modeling and predicting complex dynamical systems, particularly those involving force exchange and dissipation, is crucial for applications ranging from fluid dynamics to robotics, but presents significant challenges due to the intricate interplay of geometric constraints and energy transfer. This paper introduces Geometric Contact Flows (GFC), a novel framework leveraging Riemannian and Contact geometry as inductive biases to learn such systems. GCF constructs a latent contact Hamiltonian model encoding desirable properties like stability or energy conservation. An ensemble of contactomorphisms then adapts this model to the target dynamics while preserving these properties. This ensemble allows for uncertainty-aware geodesics that attract the system’s behavior toward the data support, enabling robust generalization and adaptation to unseen scenarios. Experiments on learning dynamics for physical systems and for controlling robots on interaction tasks demonstrate the effectiveness of our approach.

Optimization · Non-Convex

Zhangyi Liu, Feng Liu, Rui Gao, Shuang Li

We study convergence properties of the discrete-time Mean-Field Langevin Stochastic Descent-Ascent (MFL-SDA) algorithm for solving distributional minimax optimization. These problems arise in various applications, such as zero-sum games, generative adversarial networks and distributionally robust learning. Despite the significance of MFL-SDA in these contexts, the discrete-time convergence rate remains underexplored.To address this gap, we establish a last-iterate convergence rate of $O(\frac{1}{\epsilon}\log\frac{1}{\epsilon})$ for MFL-SDA. This rate is nearly optimal when compared to the complexity lower bound of its Euclidean counterpart. This rate also matches the complexity of mean-field Langevin stochastic gradient descent for distributional minimization and the outer-loop iteration complexity of an existing double-loop algorithm for distributional minimax problems.By leveraging an elementary analysis framework that avoids PDE-based techniques, we overcome previous limitations and achieve a faster convergence rate.

Optimization · Large Scale, Parallel and Distributed

Ze'ev Zukerman, Bassel Hamoud, Kfir Levy

Distributed learning methods have gained substantial momentum in recent years, with communication overhead often emerging as a critical bottleneck. Gradient compression techniques alleviate communication costs but involve an inherent trade-off between the empirical efficiency of biased compressors and the theoretical guarantees of unbiased compressors. In this work, we introduce a novel Multilevel Monte Carlo (MLMC) compression scheme that leverages biased compressors to construct statistically unbiased estimates. This approach effectively bridges the gap between biased and unbiased methods, combining the strengths of both. To showcase the versatility of our method, we apply it to popular compressors, like Top-$k$ and bit-wise compressors, resulting in enhanced variants. Furthermore, we derive an adaptive version of our approach to further improve its performance. We validate our method empirically on distributed deep learning tasks.

Applications · Computer Vision

Yinpeng Chen, Dongdong Chen, Xiyang Dai, Mengchen Liu, Yinan Feng, Youzuo Lin, Lu Yuan, Zicheng Liu

In this paper, we empirically demonstrate that natural images can be reconstructed with high fidelity from compressed representations using a simple first-order norm-plus-linear autoregressive (FINOLA) process—without relying on explicit positional information. Through systematic analysis, we observe that the learned coefficient matrices ($\mathbf{A}$ and $\mathbf{B}$) in FINOLA are typically invertible, and their product, $\mathbf{AB}^{-1}$, is diagonalizable across training runs. This structure enables a striking interpretation: FINOLA’s latent dynamics resemble a system of one-way wave equations evolving in a compressed latent space. Under this framework, each image corresponds to a unique solution of these equations. This offers a new perspective on image invariance, suggesting that the underlying structure of images may be governed by simple, invariant dynamic laws. Our findings shed light on a novel avenue for understanding and modeling visual data through the lens of latent-space dynamics and wave propagation.

General Machine Learning · Hardware and Software

Lian Liu, Xiandong Zhao, Guanchen Li, Dong Li, Wang, Yinhe Han, Xiaowei Li, ying wang

One-shot post-training pruning enhances the deployment of billion-scale large language models (LLMs), with the pruning metric playing a pivotal role in determining which weights to remove. However, existing metrics underperform due to their reliance on a simple symbolic combination of weights and activations, overlooking imbalanced weight magnitudes and the disproportionate influence of activation outliers.To overcome these limitations, we introduce \textbf{BaWA}, a novel pruning metric that systematically \underline{Ba}lances \underline{W}eight and \underline{A}ctivation distributions for more effective pruning.BaWA introduces two key innovations: \textbf{magnitude normalization}, which mitigates weight imbalance across channels for fairer pruning decisions, and \textbf{outlier regularization}, which reduces the impact of activation outliers, ensuring more appropriate channel prioritization. To further enhance its effectiveness, BaWA incorporates an efficient and automatic framework for optimizing normalization and regularization hyperparameters. Extensive experiments validate BaWA as a state-of-the-art (SOTA) pruning metric. For instance, applying BaWA to induce 2:4 sparsity in Mistral-7B reduces perplexity in language comprehension by 2.49 and improves average downstream task accuracy by 3.08\%, outperforming the previous SOTA method Wanda.

Deep Learning · Everything Else

Anshuman Chhabra, Bo Li, Jian Chen, Prasant Mohapatra, Hongfu Liu

A core data-centric learning challenge is the identification of training samples that are detrimental to model performance. Influence functions serve as a prominent tool for this task and offer a robust framework for assessing training data influence on model predictions. Despite their widespread use, their high computational cost associated with calculating the inverse of the Hessian matrix pose constraints, particularly when analyzing large-sized deep models. In this paper, we establish a bridge between identifying detrimental training samples via influence functions and outlier gradient detection. This transformation not only presents a straightforward and Hessian-free formulation but also provides insights into the role of the gradient in sample impact. Through systematic empirical evaluations, we first validate the hypothesis of our proposed outlier gradient analysis approach on synthetic datasets. We then demonstrate its effectiveness in detecting mislabeled samples in vision models and selecting data samples for improving performance of natural language processing transformer models. We also extend its use to influential sample identification for fine-tuning Large Language Models.