Gradient_Descent
-
Widening the Network Mitigates the Impact of Data Heterogeneity on FedAvg
Abstract
Federated learning (FL) enables decentralized clients to train a model collaboratively without sharing local data. A key distinction between FL and centralized learning is that clients' data are non-independent and identically distributed, which poses significant challenges in training a global model that generalizes well across heterogeneous local data distributions. In this paper, we analyze the convergence of overparameterized FedAvg with gradient descent (GD). We prove that the impact of data heterogeneity diminishes as the width of neural networks increases, ultimately vanishing when the width approaches infinity. In the infinite-width regime, we further prove that both the global and local models in FedAvg behave as linear models, and that FedAvg achieves the same generalization performance as centralized learning with the same number of GD iterations. Extensive experiments validate our theoretical findings across various network architectures, loss functions, and optimization methods.
Abstract (translated)
联邦学习（FL）允许分散的客户端在不共享本地数据的情况下协作训练模型。联邦学习与集中式学习之间的一个关键区别在于，客户端的数据是非独立且异构分布的，这给跨不同局部数据分布泛化的全局模型训练带来了重大挑战。 本文分析了使用梯度下降法（GD）进行超参数化FedAvg的收敛性。我们证明随着神经网络宽度的增加，数据异质性的负面影响逐渐减少，并在宽度趋于无穷大时完全消失。在网络宽度无限的情况下，我们进一步证明了FedAvg中的全局模型和局部模型都表现得像线性模型一样，并且当迭代次数相同时，FedAvg达到了与集中式学习相同的泛化性能。 广泛的实验验证了我们的理论发现，在各种网络架构、损失函数以及优化方法下均适用。
URL
https://arxiv.org/abs/2508.12576
PDF
-
A Fluid Antenna Enabled Physical Layer Key Generation for Next-G Wireless Networks
2025-08-16 18:39:19Jiacheng Guo, Ning Gao, Yiping Zuo, Hao Xu, Shi Jin, Kai Kit WongAbstract
As a promising physical layer security technique, physical layer key generation (PLKG) enables legitimate users to obtain secret keys from wireless channel without security infrastructures. However, in harsh propagation environments, the channel characteristic becomes unsatisfactory, the key generation rate (KGR) is significantly deteriorated. In this paper, we propose a novel fluid antenna (FA) enabled PLKG system to address this challenge. Specifically, we first derive the closed-form expression of the KGR for FA array, and then jointly optimize the precoding matrix and the antenna positions via a particle swarm optimization (PSO) algorithm. Next, to further reduce the computational complexity of the optimization procedure, we develop an alternating optimization (AO) algorithm, which combines the projected gradient descent (PGD) and the PSO. Simulation results demonstrate that by exploiting the additional spatial degree of freedom (DoF), our FA enabled PLKG system is superior to the benchmarks, such as the conventional fixed-position antenna (FPA) array and the reconfigurable intelligent surface (RIS). It is worth highlighting that compared to the conventional uniform planar antenna (UPA), the FA enabled PLKG achieves a 35.42\% KGR performance improvement under PSO algorithm and a 67.73\% KGR performance improvement under AO algorithm, respectively.
Abstract (translated)
作为一种有前景的物理层安全技术，物理层密钥生成（PLKG）使合法用户能够从无线信道中获取秘密密钥，而不依赖于任何安全基础设施。然而，在恶劣传播环境中，信道特性变得不佳，导致密钥生成率（KGR）显著下降。为此，本文提出了一种新颖的流体天线（FA）增强型PLKG系统来应对这一挑战。 具体而言，我们首先推导了FA阵列的闭式KGR表达式，并通过粒子群优化算法联合优化预编码矩阵和天线位置。接下来，为了进一步降低优化过程中的计算复杂度，我们开发了一种交替优化（AO）算法，该算法结合了投影梯度下降法（PGD）和粒子群优化算法。 仿真结果表明，在利用额外的空间自由度后，我们的FA增强型PLKG系统在性能上优于基准方案，如传统的固定位置天线（FPA）阵列以及可重构智能表面（RIS）。尤其值得注意的是，与传统均匀平面阵列（UPA）相比，基于FA的PLKG在使用PSO算法时，KGR性能提升了35.42%，而在采用AO算法时，则达到了67.73%的提升。
URL
https://arxiv.org/abs/2509.00018
PDF
-
Implicit Hypergraph Neural Network
2025-08-16 16:58:59Akash Choudhuri, Yongjian Zhong, Bijaya AdhikariAbstract
Hypergraphs offer a generalized framework for capturing high-order relationships between entities and have been widely applied in various domains, including healthcare, social networks, and bioinformatics. Hypergraph neural networks, which rely on message-passing between nodes over hyperedges to learn latent representations, have emerged as the method of choice for predictive tasks in many of these domains. These approaches typically perform only a small number of message-passing rounds to learn the representations, which they then utilize for predictions. The small number of message-passing rounds comes at a cost, as the representations only capture local information and forego long-range high-order dependencies. However, as we demonstrate, blindly increasing the message-passing rounds to capture long-range dependency also degrades the performance of hyper-graph neural networks. Recent works have demonstrated that implicit graph neural networks capture long-range dependencies in standard graphs while maintaining performance. Despite their popularity, prior work has not studied long-range dependency issues on hypergraph neural networks. Here, we first demonstrate that existing hypergraph neural networks lose predictive power when aggregating more information to capture long-range dependency. We then propose Implicit Hypergraph Neural Network (IHNN), a novel framework that jointly learns fixed-point representations for both nodes and hyperedges in an end-to-end manner to alleviate this issue. Leveraging implicit differentiation, we introduce a tractable projected gradient descent approach to train the model efficiently. Extensive experiments on real-world hypergraphs for node classification demonstrate that IHNN outperforms the closest prior works in most settings, establishing a new state-of-the-art in hypergraph learning.
Abstract (translated)
超图提供了一种捕捉实体之间高阶关系的通用框架，并在医疗保健、社交网络和生物信息学等各个领域得到了广泛应用。基于超边进行节点间消息传递以学习潜在表示的超图神经网络，已成为这些领域许多预测任务中的首选方法。这些方法通常只执行少量的消息传递轮次来学习表示，然后利用这些表示来进行预测。这种少量的消息传递轮次会导致表示仅捕获局部信息，并忽略长程高阶依赖关系。然而，我们证明盲目增加消息传递轮数以捕捉长程依赖同样会降低超图神经网络的性能。 近期工作表明，隐式图神经网络能够在标准图中捕获长距离依赖性的同时保持良好的性能表现，但之前的工作没有研究过超图神经网络中的远程依赖问题。在这里，我们首先证明现有的超图神经网络在聚合更多信息以捕捉长程依赖时会失去预测能力。然后，我们提出了一种新的框架——隐式超图神经网络（IHNN），该框架通过端到端的方式共同学习节点和超边的固定点表示，从而缓解这一问题。利用隐式微分，我们引入了一种可操作的投影梯度下降方法来高效训练模型。 在真实世界的超图上进行节点分类任务的广泛实验表明，在大多数情况下，IHNN超越了最近的相关工作，建立了新的最先进的技术成果。
URL
https://arxiv.org/abs/2508.14101
PDF
-
Provable In-Context Vector Arithmetic via Retrieving Task Concepts
2025-08-13 13:54:44Dake Bu, Wei Huang, Andi Han, Atsushi Nitanda, Qingfu Zhang, Hau-San Wong, Taiji SuzukiAbstract
In-context learning (ICL) has garnered significant attention for its ability to grasp functions/tasks from demonstrations. Recent studies suggest the presence of a latent task/function vector in LLMs during ICL. Merullo et al. (2024) showed that LLMs leverage this vector alongside the residual stream for Word2Vec-like vector arithmetic, solving factual-recall ICL tasks. Additionally, recent work empirically highlighted the key role of Question-Answer data in enhancing factual-recall capabilities. Despite these insights, a theoretical explanation remains elusive. To move one step forward, we propose a theoretical framework building on empirically grounded hierarchical concept modeling. We develop an optimization theory, showing how nonlinear residual transformers trained via gradient descent on cross-entropy loss perform factual-recall ICL tasks via vector arithmetic. We prove 0-1 loss convergence and show the strong generalization, including robustness to concept recombination and distribution shifts. These results elucidate the advantages of transformers over static embedding predecessors. Empirical simulations corroborate our theoretical insights.
Abstract (translated)
在上下文学习（ICL）中，模型从演示中掌握功能/任务的能力引起了广泛关注。近期研究表明，在进行ICL时，大型语言模型（LLM）内部存在一个潜在的任务/函数向量。Merullo等人(2024)发现，LLMs利用这一向量与残差流结合来进行类似Word2Vec的向量运算，从而完成事实回忆型ICL任务。此外，近期的工作通过实证研究强调了问答数据在提升事实回忆能力方面的重要作用。尽管有了这些见解，但理论解释仍然缺失。 为了向前推进一步，我们提出了一种基于经验基础分层概念建模的理论框架。我们开发了一个优化理论，展示了非线性残差变压器如何通过对交叉熵损失进行梯度下降训练来通过向量运算执行事实回忆型ICL任务。我们证明了0-1损失收敛，并表明其具有强大的泛化能力，包括对概念重组和分布变化的鲁棒性。这些结果阐明了变压器相对于静态嵌入前体的优势。 实证模拟验证了我们的理论见解。
URL
https://arxiv.org/abs/2508.09820
PDF
-
Physics-guided Deep Unfolding Network for Enhanced Kronecker Compressive sensing
2025-08-13 06:21:58Gang Qu, Ping Wang, Siming Zheng, Xin YuanAbstract
Deep networks have achieved remarkable success in image compressed sensing (CS) task, namely reconstructing a high-fidelity image from its compressed measurement. However, existing works are deficient inincoherent compressed measurement at sensing phase and implicit measurement representations at reconstruction phase, limiting the overall performance. In this work, we answer two questions: 1) how to improve the measurement incoherence for decreasing the ill-posedness; 2) how to learn informative representations from measurements. To this end, we propose a novel asymmetric Kronecker CS (AKCS) model and theoretically present its better incoherence than previous Kronecker CS with minimal complexity increase. Moreover, we reveal that the unfolding networks' superiority over non-unfolding ones result from sufficient gradient descents, called explicit measurement representations. We propose a measurement-aware cross attention (MACA) mechanism to learn implicit measurement representations. We integrate AKCS and MACA into widely-used unfolding architecture to get a measurement-enhanced unfolding network (MEUNet). Extensive experiences demonstrate that our MEUNet achieves state-of-the-art performance in reconstruction accuracy and inference speed.
Abstract (translated)
深度网络在图像压缩感知（CS）任务中取得了显著的成功，即从其压缩测量值重构出高质量的图像。然而，现有的研究工作在传感阶段缺乏不相干的压缩测量，在重建阶段也未能明确表示测量结果，这限制了整体性能的表现。本文旨在回答两个问题：1）如何改进测量的不相关性以减少不适定性；2）如何从测量数据中学习出具有信息量的表示。 为此，我们提出了一个新颖的非对称克罗内克压缩感知（AKCS）模型，并从理论上证明了其比之前的克罗内克压缩感知方法在复杂度最小增加的情况下具备更好的不相关性。此外，我们揭示了解折叠网络优于非解折叠网络的原因在于充分的梯度下降过程，即显式的测量表示。 为了学习隐含的测量表示，我们提出了一个基于测量的认知交叉注意力（MACA）机制。我们将AKCS和MACA整合到广泛使用的解折叠架构中，得到了一种增强测量信息的解折叠网络（MEUNet）。大量的实验经验表明，我们的MEUNet在重建精度和推理速度方面达到了最先进的性能水平。
URL
https://arxiv.org/abs/2508.09528
PDF
-
GreenTEA: Gradient Descent with Topic-modeling and Evolutionary Auto-prompting
2025-08-12 06:48:30Zheng Dong, Luming Shang, Gabriela OlintoAbstract
High-quality prompts are crucial for Large Language Models (LLMs) to achieve exceptional performance. However, manually crafting effective prompts is labor-intensive and demands significant domain expertise, limiting its scalability. Existing automatic prompt optimization methods either extensively explore new prompt candidates, incurring high computational costs due to inefficient searches within a large solution space, or overly exploit feedback on existing prompts, risking suboptimal optimization because of the complex prompt landscape. To address these challenges, we introduce GreenTEA, an agentic LLM workflow for automatic prompt optimization that balances candidate exploration and knowledge exploitation. It leverages a collaborative team of agents to iteratively refine prompts based on feedback from error samples. An analyzing agent identifies common error patterns resulting from the current prompt via topic modeling, and a generation agent revises the prompt to directly address these key deficiencies. This refinement process is guided by a genetic algorithm framework, which simulates natural selection by evolving candidate prompts through operations such as crossover and mutation to progressively optimize model performance. Extensive numerical experiments conducted on public benchmark datasets suggest the superior performance of GreenTEA against human-engineered prompts and existing state-of-the-arts for automatic prompt optimization, covering logical and quantitative reasoning, commonsense, and ethical decision-making.
Abstract (translated)
高质量的提示对于大型语言模型（LLM）实现卓越性能至关重要。然而，手动设计有效的提示既费时又需要深厚的专业知识，这限制了其可扩展性。现有的自动提示优化方法要么广泛探索新的提示候选方案，在庞大的解决方案空间中进行低效搜索导致计算成本高昂；要么过度利用现有提示的反馈，由于复杂的提示景观而面临次优优化的风险。为了解决这些挑战，我们引入了一种名为GreenTEA的代理LLM工作流程，用于自动提示优化，它在候选探索和知识利用之间取得了平衡。 GreenTEA通过一个协作的智能体团队根据错误样本的反馈迭代地改进提示。其中，分析智能体通过主题建模识别当前提示产生的常见错误模式，而生成智能体则修订提示以直接解决这些关键缺陷。该优化过程由遗传算法框架引导，模拟自然选择，在交叉和变异等操作下进化候选提示，逐步优化模型性能。 在公共基准数据集上进行的大量数值实验表明，GreenTEA相对于人为设计的提示以及现有的自动提示优化方法中的最新技术具有更优越的表现，涵盖了逻辑和定量推理、常识及伦理决策等领域。
URL
https://arxiv.org/abs/2508.16603
PDF
-
Fast weight programming and linear transformers: from machine learning to neurobiology
2025-08-11 19:50:03Kazuki Irie, Samuel J. GershmanAbstract
Recent advances in artificial neural networks for machine learning, and language modeling in particular, have established a family of recurrent neural network (RNN) architectures that, unlike conventional RNNs with vector-form hidden states, use two-dimensional (2D) matrix-form hidden states. Such 2D-state RNNs, known as Fast Weight Programmers (FWPs), can be interpreted as a neural network whose synaptic weights (called fast weights) dynamically change over time as a function of input observations, and serve as short-term memory storage; corresponding synaptic weight modifications are controlled or programmed by another network (the programmer) whose parameters are trained (e.g., by gradient descent). In this Primer, we review the technical foundations of FWPs, their computational characteristics, and their connections to transformers and state space models. We also discuss connections between FWPs and models of synaptic plasticity in the brain, suggesting a convergence of natural and artificial intelligence.
Abstract (translated)
最近在机器学习领域，特别是在语言模型方面的进展，确立了一类循环神经网络（RNN）架构。与传统的使用向量形式隐状态的RNN不同，这类新架构采用二维矩阵形式的隐藏状态。这种具有2D状态的RNN被称为快速权重编程器（FWPs），可以被解释为一个其突触权重（称为快权重）会随输入观察结果动态变化的神经网络，并且这些权重作为短期记忆存储；相应的突触权重调整由另一个网络（即编程器）控制，该网络的参数通过训练获得（例如，通过梯度下降法）。在本文中，我们回顾了FWPs的技术基础、计算特性以及它们与变压器和状态空间模型之间的联系。此外，我们还讨论了FWPs与大脑突触可塑性模型之间的关联，这表明自然智能和人工智能正在趋于融合。
URL
https://arxiv.org/abs/2508.08435
PDF
-
Multi-head Transformers Provably Learn Symbolic Multi-step Reasoning via Gradient Descent
2025-08-11 17:40:47Tong Yang, Yu Huang, Yingbin Liang, Yuejie ChiAbstract
Transformers have demonstrated remarkable capabilities in multi-step reasoning tasks. However, understandings of the underlying mechanisms by which they acquire these abilities through training remain limited, particularly from a theoretical standpoint. This work investigates how transformers learn to solve symbolic multi-step reasoning problems through chain-of-thought processes, focusing on path-finding in trees. We analyze two intertwined tasks: a backward reasoning task, where the model outputs a path from a goal node to the root, and a more complex forward reasoning task, where the model implements two-stage reasoning by first identifying the goal-to-root path and then reversing it to produce the root-to-goal path. Our theoretical analysis, grounded in the dynamics of gradient descent, shows that trained one-layer transformers can provably solve both tasks with generalization guarantees to unseen trees. In particular, our multi-phase training dynamics for forward reasoning elucidate how different attention heads learn to specialize and coordinate autonomously to solve the two subtasks in a single autoregressive path. These results provide a mechanistic explanation of how trained transformers can implement sequential algorithmic procedures. Moreover, they offer insights into the emergence of reasoning abilities, suggesting that when tasks are structured to take intermediate chain-of-thought steps, even shallow multi-head transformers can effectively solve problems that would otherwise require deeper architectures.
Abstract (translated)
Transformer模型在多步推理任务中表现出色，但它们通过训练获得这些能力的内在机制，尤其是在理论层面上的理解仍然有限。本研究探讨了Transformer是如何通过链式思维过程学习解决符号化多步推理问题的，特别是关注于树形结构中的路径寻找。我们分析了两个相互交织的任务：一个是反向推理任务，在此任务中模型输出从目标节点到根节点的路径；另一个是更复杂的正向推理任务，即模型首先识别出从目标节点到根节点的路径，然后反转这条路径以生成从根节点到目标节点的路径。基于梯度下降动力学的理论分析表明，经过训练的一层Transformer可以证明解决这两个任务，并且对未见过的树形结构具有泛化保证。 特别地，我们为正向推理开发出一个多阶段训练动态过程，阐明了不同的注意力头如何自主学习专业化并协作以在单一自回归路径中解决两个子任务。这些结果提供了关于经过训练的Transformer如何实现顺序算法程序的具体机制解释。此外，它们还为推理能力的出现提供了见解：当任务结构被设计成需要中间链式思维步骤时，即使是浅层多头Transformer也能有效解决问题，而这些问题在其他情况下可能需要更深层的架构来解决。
URL
https://arxiv.org/abs/2508.08222
PDF
-
SGD Convergence under Stepsize Shrinkage in Low-Precision Training
2025-08-10 02:25:48Vincent-Daniel YunAbstract
Low-precision training has become essential for reducing the computational and memory costs of large-scale deep learning. However, quantization of gradients introduces both magnitude shrinkage and additive noise, which can alter the convergence behavior of stochastic gradient descent (SGD). In this work, we study the convergence of SGD under a gradient shrinkage model, where each stochastic gradient is scaled by a factor $q_k \in (0,1]$ and perturbed by zero-mean quantization noise. We show that this shrinkage is equivalent to replacing the nominal stepsize $\mu_k$ with an effective stepsize $\mu_k q_k$, which slows convergence when $q_{\min} < 1$. Under standard smoothness and bounded-variance assumptions, we prove that low-precision SGD still converges, but at a reduced rate determined by $q_{\min}$, and with an increased asymptotic error floor due to quantization noise. We theoretically analyze how reduced numerical precision slows down training by modeling it as gradient shrinkage in the standard SGD convergence framework.
Abstract (translated)
低精度训练已成为减少大规模深度学习计算和内存成本的关键方法。然而，梯度量化会引入幅度衰减以及加性噪声，这些都可能改变随机梯度下降（SGD）的收敛行为。在这项工作中，我们研究了在一种梯度缩放模型下的SGD收敛性，在这种模型中，每个随机梯度被一个因子$q_k \in (0,1]$ 缩小，并且受到零均值量化噪声的影响。 我们的研究表明，该缩放等同于将名义学习率$\mu_k$替换为有效学习率$\mu_k q_k$。当最小的缩放因子$q_{\min} < 1$时，这会导致收敛速度变慢。在标准平滑性和有界方差假设下，我们证明了低精度SGD仍然能够收敛，但以一个由$q_{\min}$决定的速度减缓，并且由于量化噪声导致最终误差增加。 从理论上讲，我们通过将数值精度降低建模为标准SGD收敛框架中的梯度缩放来分析其如何减慢训练速度。
URL
https://arxiv.org/abs/2508.07142
PDF
-
FIVA: Federated Inverse Variance Averaging for Universal CT Segmentation with Uncertainty Estimation
2025-08-08 11:34:01Asim Ukaye, Numan Saeed, Karthik NandakumarAbstract
Different CT segmentation datasets are typically obtained from different scanners under different capture settings and often provide segmentation labels for a limited and often disjoint set of organs. Using these heterogeneous data effectively while preserving patient privacy can be challenging. This work presents a novel federated learning approach to achieve universal segmentation across diverse abdominal CT datasets by utilizing model uncertainty for aggregation and predictive uncertainty for inference. Our approach leverages the inherent noise in stochastic mini-batch gradient descent to estimate a distribution over the model weights to provide an on-the-go uncertainty over the model parameters at the client level. The parameters are then aggregated at the server using the additional uncertainty information using a Bayesian-inspired inverse-variance aggregation scheme. Furthermore, the proposed method quantifies prediction uncertainty by propagating the uncertainty from the model weights, providing confidence measures essential for clinical decision-making. In line with recent work shown, predictive uncertainty is utilized in the inference stage to improve predictive performance. Experimental evaluations demonstrate the effectiveness of this approach in improving both the quality of federated aggregation and uncertainty-weighted inference compared to previously established baselines. The code for this work is made available at: this https URL
Abstract (translated)
不同的CT分割数据集通常来自不同类型的扫描仪和不同的采集设置，并且往往只为有限的、不连续的一组器官提供分割标签。有效地使用这些异质性数据并同时保护患者隐私是一项挑战。这项工作提出了一种新颖的联邦学习方法，通过利用模型不确定性和预测不确定性来实现跨多样腹部CT数据集的通用分割。我们的方法利用随机小批量梯度下降中的固有噪声来估计模型权重上的分布，从而在客户端级别实时提供关于模型参数的不确定性。然后，在服务器端使用基于贝叶斯的思想和逆方差聚合方案来整合这些参数，并结合额外的不确定性信息进行汇总。 此外，所提出的方法通过从模型权重中传播不确定性来量化预测中的不确定性和置信度量，这对临床决策至关重要。正如最近的工作所示，我们在推理阶段利用了预测的不确定性以提高预测性能。实验评估表明，与先前建立的基础方法相比，该方法在改善联邦聚合质量和基于不确定性的加权推断方面是有效的。 此工作的代码可在以下网址获取：this https URL
URL
https://arxiv.org/abs/2508.09196
PDF
-
Neural Field Representations of Mobile Computational Photography
Abstract
Over the past two decades, mobile imaging has experienced a profound transformation, with cell phones rapidly eclipsing all other forms of digital photography in popularity. Today's cell phones are equipped with a diverse range of imaging technologies - laser depth ranging, multi-focal camera arrays, and split-pixel sensors - alongside non-visual sensors such as gyroscopes, accelerometers, and magnetometers. This, combined with on-board integrated chips for image and signal processing, makes the cell phone a versatile pocket-sized computational imaging platform. Parallel to this, we have seen in recent years how neural fields - small neural networks trained to map continuous spatial input coordinates to output signals - enable the reconstruction of complex scenes without explicit data representations such as pixel arrays or point clouds. In this thesis, I demonstrate how carefully designed neural field models can compactly represent complex geometry and lighting effects. Enabling applications such as depth estimation, layer separation, and image stitching directly from collected in-the-wild mobile photography data. These methods outperform state-of-the-art approaches without relying on complex pre-processing steps, labeled ground truth data, or machine learning priors. Instead, they leverage well-constructed, self-regularized models that tackle challenging inverse problems through stochastic gradient descent, fitting directly to raw measurements from a smartphone.
Abstract (translated)
在过去二十年里，移动成像经历了深刻的变革，手机迅速超越了所有其他形式的数字摄影，在流行度上占据主导地位。今天的手机配备了各种成像技术——包括激光深度测距、多焦点相机阵列和分隔像素传感器——以及非视觉传感器，如陀螺仪、加速度计和磁力计。结合内置的图像和信号处理集成电路芯片，使手机成为一种多功能的手持式计算成像平台。 与此同时，在近年来我们见证了神经场（小型经过训练以将连续空间输入坐标映射到输出信号的小型神经网络）的发展，它们能够无需显式的数据表示（如像素阵列或点云）来重建复杂场景。在我的论文中，我展示了精心设计的神经场模型如何紧凑地代表复杂的几何结构和光照效果，并支持从收集的实际移动摄影数据直接进行深度估计、图层分离和图像拼接等应用。 这些方法在不依赖复杂的预处理步骤、标记的真实数据或机器学习先验的情况下超越了现有最佳的方法。相反，它们利用精心构造的、自我调节的模型通过随机梯度下降直接拟合智能手机采集的原始测量值来解决具有挑战性的逆向问题。
URL
https://arxiv.org/abs/2508.05907
PDF
-
Comparative Analysis of Novel NIRMAL Optimizer Against Adam and SGD with Momentum
2025-08-06 10:30:22Nirmal Gaud, Surej Mouli, Preeti Katiyar, Vaduguru Venkata RamyaAbstract
This study proposes NIRMAL (Novel Integrated Robust Multi-Adaptation Learning), a novel optimization algorithm that combines multiple strategies inspired by the movements of the chess piece. These strategies include gradient descent, momentum, stochastic perturbations, adaptive learning rates, and non-linear transformations. We carefully evaluated NIRMAL against two widely used and successful optimizers, Adam and SGD with Momentum, on four benchmark image classification datasets: MNIST, FashionMNIST, CIFAR-10, and CIFAR-100. The custom convolutional neural network (CNN) architecture is applied on each dataset. The experimental results show that NIRMAL achieves competitive performance, particularly on the more challenging CIFAR-100 dataset, where it achieved a test accuracy of 45.32\%and a weighted F1-score of 0.4328. This performance surpasses Adam (41.79\% accuracy, 0.3964 F1-score) and closely matches SGD with Momentum (46.97\% accuracy, 0.4531 F1-score). Also, NIRMAL exhibits robust convergence and strong generalization capabilities, especially on complex datasets, as evidenced by stable training results in loss and accuracy curves. These findings underscore NIRMAL's significant ability as a versatile and effective optimizer for various deep learning tasks.
Abstract (translated)
这项研究提出了NIRMAL（新颖的综合鲁棒多适应学习），这是一种新的优化算法，结合了多种策略，这些策略受到了国际象棋棋子移动方式的启发。这些策略包括梯度下降、动量、随机扰动、自适应学习率和非线性变换。我们仔细地将NIRMAL与两种广泛使用且成功的优化器——Adam和带动量的SGD，在四个基准图像分类数据集上进行了比较：MNIST、FashionMNIST、CIFAR-10 和 CIFAR-100。针对每个数据集，应用了自定义的卷积神经网络（CNN）架构。实验结果表明，NIRMAL实现了具有竞争力的表现，特别是在更具挑战性的CIFAR-100 数据集中，它的测试准确率为45.32%，加权F1分数为0.4328。这一性能超过了Adam（准确率 41.79%、F1 分数 0.3964）并且接近于带动量的SGD（准确率 46.97%、F1 分数 0.4531）。此外，NIRMAL展示了鲁棒收敛和强大的泛化能力，尤其是在复杂数据集上，这一点通过损失和精度曲线中的稳定训练结果得到了证实。这些发现强调了 NIRMAL 在各种深度学习任务中作为一种灵活且有效的优化器的重要性。
URL
https://arxiv.org/abs/2508.04293
PDF
-
When Good Sounds Go Adversarial: Jailbreaking Audio-Language Models with Benign Inputs
2025-08-05 12:14:01Bodam Kim, Hiskias Dingeto, Taeyoun Kwon, Dasol Choi, DongGeon Lee, Haon Park, JaeHoon Lee, Jongho ShinAbstract
As large language models become increasingly integrated into daily life, audio has emerged as a key interface for human-AI interaction. However, this convenience also introduces new vulnerabilities, making audio a potential attack surface for adversaries. Our research introduces WhisperInject, a two-stage adversarial audio attack framework that can manipulate state-of-the-art audio language models to generate harmful content. Our method uses imperceptible perturbations in audio inputs that remain benign to human listeners. The first stage uses a novel reward-based optimization method, Reinforcement Learning with Projected Gradient Descent (RL-PGD), to guide the target model to circumvent its own safety protocols and generate harmful native responses. This native harmful response then serves as the target for Stage 2, Payload Injection, where we use Projected Gradient Descent (PGD) to optimize subtle perturbations that are embedded into benign audio carriers, such as weather queries or greeting messages. Validated under the rigorous StrongREJECT, LlamaGuard, as well as Human Evaluation safety evaluation framework, our experiments demonstrate a success rate exceeding 86% across Qwen2.5-Omni-3B, Qwen2.5-Omni-7B, and Phi-4-Multimodal. Our work demonstrates a new class of practical, audio-native threats, moving beyond theoretical exploits to reveal a feasible and covert method for manipulating AI behavior.
Abstract (translated)
随着大型语言模型在日常生活中越来越普及，音频已成为人机交互的关键接口。然而，这种便利性也带来了新的安全漏洞，使音频成为潜在的攻击面。我们的研究介绍了WhisperInject，这是一种两阶段的对抗性音频攻击框架，可以操纵最先进的音频语言模型生成有害内容。我们使用了微小的音频输入扰动，这些扰动对人类听者来说是无害且不易察觉的。 第一阶段采用了一种基于奖励的优化方法——强化学习与投影梯度下降（RL-PGD），引导目标模型绕过自身的安全协议并生成有害响应。这种有害响应随后成为第二阶段“载荷注入”的攻击目标，在此阶段中，我们使用投影梯度下降（PGD）来优化嵌入到无害音频载体中的细微扰动，例如天气查询或问候信息。 在严格的StrongREJECT、LlamaGuard以及人类评估安全框架下验证后，我们的实验结果显示，WhisperInject在Qwen2.5-Omni-3B、Qwen2.5-Omni-7B和Phi-4-Multimodal模型上的成功率达到超过86%。这项工作揭示了一类新的实用且原生的音频威胁，并从理论性攻击方法转向了实际可行且隐蔽地操纵AI行为的方法。
URL
https://arxiv.org/abs/2508.03365
PDF
-
Adversarial Attacks on Reinforcement Learning-based Medical Questionnaire Systems: Input-level Perturbation Strategies and Medical Constraint Validation
2025-08-05 11:10:49Peizhuo LiuAbstract
RL-based medical questionnaire systems have shown great potential in medical scenarios. However, their safety and robustness remain unresolved. This study performs a comprehensive evaluation on adversarial attack methods to identify and analyze their potential vulnerabilities. We formulate the diagnosis process as a Markov Decision Process (MDP), where the state is the patient responses and unasked questions, and the action is either to ask a question or to make a diagnosis. We implemented six prevailing major attack methods, including the Fast Gradient Signed Method (FGSM), Projected Gradient Descent (PGD), Carlini & Wagner Attack (C&W) attack, Basic Iterative Method (BIM), DeepFool, and AutoAttack, with seven epsilon values each. To ensure the generated adversarial examples remain clinically plausible, we developed a comprehensive medical validation framework consisting of 247 medical constraints, including physiological bounds, symptom correlations, and conditional medical constraints. We achieved a 97.6% success rate in generating clinically plausible adversarial samples. We performed our experiment on the National Health Interview Survey (NHIS) dataset (this https URL), which consists of 182,630 samples, to predict the participant's 4-year mortality rate. We evaluated our attacks on the AdaptiveFS framework proposed in arXiv:2004.00994. Our results show that adversarial attacks could significantly impact the diagnostic accuracy, with attack success rates ranging from 33.08% (FGSM) to 64.70% (AutoAttack). Our work has demonstrated that even under strict medical constraints on the input, such RL-based medical questionnaire systems still show significant vulnerabilities.
Abstract (translated)
基于强化学习（RL）的医疗问卷系统在医学场景中展现了巨大的潜力，但其安全性和鲁棒性问题尚未解决。本研究对对抗攻击方法进行了全面评估，以识别和分析潜在的安全漏洞。我们把诊断过程描述为一个马尔可夫决策过程（MDP），其中状态是患者的回答以及未问的问题，动作则是提问或做出诊断。 我们在六种主要的攻击方法上进行实施：快速梯度符号法（FGSM）、投影梯度下降（PGD）、Carlini & Wagner 攻击（C&W）攻击、基本迭代方法（BIM）、DeepFool 和 AutoAttack。每种方法都使用了七个不同的ε值。 为了确保生成的对抗样本在临床上合理，我们建立了一个包含247项医学限制条件的全面验证框架，包括生理界限、症状相关性和条件性医疗限制。我们实现了97.6%的成功率，使得生成的对抗样本保持临床合理性。 我们的实验使用了全国健康访谈调查（NHIS）数据集进行，该数据集包含了182,630个样本，并预测参与者的四年死亡率。我们在提出的AdaptiveFS框架上评估了我们的攻击方法，发现这些攻击可以显著影响诊断准确性，成功率从FGSM的33.08%到AutoAttack的64.70%不等。 这项研究表明，即使在输入数据上有严格的医学限制条件下，基于强化学习的医疗问卷系统仍然存在明显的安全漏洞。
URL
https://arxiv.org/abs/2508.05677
PDF
-
How Diffusion Prior Landscapes Shape the Posterior in Blind Deconvolution
2025-08-04 21:53:12Minh-Hai Nguyen, Edouard Pauwels, Pierre WeissAbstract
The Maximum A Posteriori (MAP) estimation is a widely used framework in blind deconvolution to recover sharp images from blurred observations. The estimated image and blur filter are defined as the maximizer of the posterior distribution. However, when paired with sparsity-promoting image priors, MAP estimation has been shown to favors blurry solutions, limiting its effectiveness. In this paper, we revisit this result using diffusion-based priors, a class of models that capture realistic image distributions. Through an empirical examination of the prior's likelihood landscape, we uncover two key properties: first, blurry images tend to have higher likelihoods; second, the landscape contains numerous local minimizers that correspond to natural images. Building on these insights, we provide a theoretical analysis of the blind deblurring posterior. This reveals that the MAP estimator tends to produce sharp filters (close to the Dirac delta function) and blurry solutions. However local minimizers of the posterior, which can be obtained with gradient descent, correspond to realistic, natural images, effectively solving the blind deconvolution problem. Our findings suggest that overcoming MAP's limitations requires good local initialization to local minima in the posterior landscape. We validate our analysis with numerical experiments, demonstrating the practical implications of our insights for designing improved priors and optimization techniques.
Abstract (translated)
最大后验估计（MAP）是盲去卷积中广泛使用的一种框架，用于从模糊观测中恢复清晰图像。在该框架下，估计的图像和模糊滤波器被定义为后验分布的最大值。然而，当与促进图像稀疏性的先验条件结合时，MAP估计会倾向于产生模糊解，从而限制了其有效性。 本文重新审视了这一结果，并使用基于扩散的先验模型进行分析。这是一种捕捉现实图像分布的模型类型。通过实证检查先验似然景观，我们发现了两个关键特性：首先，模糊图像往往具有更高的可能性；其次，该景观包含大量局部极小值点，这些极小值对应于自然图像。 基于这些见解，我们对盲去卷积后验进行了理论分析。这揭示了MAP估计器倾向于产生锐利的滤波器（接近狄拉克δ函数）和模糊解。然而，通过梯度下降等方法可以获取到后验中的局部极小值点，这些局部极小值对应于真实的自然图像，从而有效解决了盲去卷积问题。 我们的研究结果表明，克服MAP局限性的关键是找到一个好的局部初始化以到达后验景观中的局部最小值点。我们通过数值实验验证了我们的分析，并展示了我们的见解对设计改进的先验条件和优化技术的实际意义。
URL
https://arxiv.org/abs/2508.02923
PDF
-
Towards Real Unsupervised Anomaly Detection Via Confident Meta-Learning
2025-08-04 11:03:12Muhammad Aqeel, Shakiba Sharifi, Marco Cristani, Francesco SettiAbstract
So-called unsupervised anomaly detection is better described as semi-supervised, as it assumes all training data are nominal. This assumption simplifies training but requires manual data curation, introducing bias and limiting adaptability. We propose Confident Meta-learning (CoMet), a novel training strategy that enables deep anomaly detection models to learn from uncurated datasets where nominal and anomalous samples coexist, eliminating the need for explicit filtering. Our approach integrates Soft Confident Learning, which assigns lower weights to low-confidence samples, and Meta-Learning, which stabilizes training by regularizing updates based on training validation loss covariance. This prevents overfitting and enhances robustness to noisy data. CoMet is model-agnostic and can be applied to any anomaly detection method trainable via gradient descent. Experiments on MVTec-AD, VIADUCT, and KSDD2 with two state-of-the-art models demonstrate the effectiveness of our approach, consistently improving over the baseline methods, remaining insensitive to anomalies in the training set, and setting a new state-of-the-art across all datasets.
Abstract (translated)
所谓的无监督异常检测实际上更接近于半监督学习，因为它假设所有训练数据都是正常的。这种假设虽然简化了模型的训练过程，但需要手动进行数据整理以去除或标注异常样本，这不仅引入了偏差，还限制了模型适应新数据的能力。我们提出了一种新的训练策略——自信元学习（Confident Meta-learning, CoMet），这种方法可以使深度异常检测模型从未经整理的数据集中学习，在这些数据集中正常和异常样本共存。CoMet消除了显式过滤的需要，使得模型能够更加灵活地处理包含多种类型样本的真实世界数据。 我们的方法结合了软自信学习（Soft Confident Learning）技术，该技术对置信度较低的样本赋予较小的权重，并使用元学习（Meta-Learning）来通过基于训练验证损失协方差的正则化更新规则稳定模型训练。这种方法不仅能够防止过拟合现象，还能增强模型在处理嘈杂数据时的鲁棒性。 CoMet算法与特定的模型无关，可以应用于任何可以通过梯度下降进行训练的异常检测方法中。我们通过在MVTec-AD、VIADUCT和KSDD2数据集上使用两种最先进的模型进行了实验验证，结果表明我们的方法能够显著优于基准方法，并且不受训练集中存在的异常样本的影响，在所有测试的数据集中均取得了新的最优性能。
URL
https://arxiv.org/abs/2508.02293
PDF
-
Categorical Construction of Logically Verifiable Neural Architectures
2025-08-02 04:30:05Logan NyeAbstract
Neural networks excel at pattern recognition but struggle with reliable logical reasoning, often violating basic logical principles during inference. We address this limitation by developing a categorical framework that systematically constructs neural architectures with provable logical guarantees. Our approach treats logical theories as algebraic structures called Lawvere theories, which we transform into neural networks using categorical algebra in the 2-category of parametric maps. Unlike existing methods that impose logical constraints during training, our categorical construction embeds logical principles directly into the network's architectural structure, making logical violations mathematically impossible. We demonstrate this framework by constructing differentiable neural architectures for propositional logic that preserve boolean reasoning while remaining trainable via gradient descent. Our main theoretical result establishes a bijective correspondence between finitary logical theories and neural architectures, proving that every logically constrained network arises uniquely from our construction. This extends Categorical Deep Learning beyond geometric symmetries to semantic constraints, enabling automatic derivation of verified architectures from logical specifications. The framework provides mathematical foundations for trustworthy AI systems, with applications to theorem proving, formal verification, and safety-critical reasoning tasks requiring verifiable logical behavior.
Abstract (translated)
神经网络在模式识别方面表现出色，但在可靠的逻辑推理方面存在困难，在推理过程中常常违反基本的逻辑原则。我们通过开发一个范畴框架来解决这一限制，该框架系统地构建具有可证明逻辑保证的神经架构。我们的方法将逻辑理论视为称为Lawvere理论的代数结构，并使用参数映射2-范畴中的范畴代数将其转换为神经网络。与现有方法在训练期间施加逻辑约束不同，我们的范畴构造直接将逻辑原则嵌入到网络的架构中，使得逻辑违规从数学上是不可能的。我们通过构建用于命题逻辑的不同iable神经架构来展示这一框架，这些架构在保持布尔推理的同时可以通过梯度下降进行训练。我们主要的理论成果确立了有限逻辑理论和神经架构之间的一一对应关系，证明每个受逻辑约束的网络都唯一地源自我们的构造。这将范畴深度学习从几何对称性扩展到语义约束，使能够自动从逻辑规范中推导出可验证的架构。该框架为值得信赖的人工智能系统提供了数学基础，并在定理证明、形式验证和需要可验证逻辑行为的安全关键推理任务中有着应用。
URL
https://arxiv.org/abs/2508.11647
PDF
-
TITAN-Guide: Taming Inference-Time AligNment for Guided Text-to-Video Diffusion Models
2025-08-01 03:26:18Christian Simon, Masato Ishii, Akio Hayakawa, Zhi Zhong, Shusuke Takahashi, Takashi Shibuya, Yuki MitsufujiAbstract
In the recent development of conditional diffusion models still require heavy supervised fine-tuning for performing control on a category of tasks. Training-free conditioning via guidance with off-the-shelf models is a favorable alternative to avoid further fine-tuning on the base model. However, the existing training-free guidance frameworks either have heavy memory requirements or offer sub-optimal control due to rough estimation. These shortcomings limit the applicability to control diffusion models that require intense computation, such as Text-to-Video (T2V) diffusion models. In this work, we propose Taming Inference Time Alignment for Guided Text-to-Video Diffusion Model, so-called TITAN-Guide, which overcomes memory space issues, and provides more optimal control in the guidance process compared to the counterparts. In particular, we develop an efficient method for optimizing diffusion latents without backpropagation from a discriminative guiding model. In particular, we study forward gradient descents for guided diffusion tasks with various options on directional directives. In our experiments, we demonstrate the effectiveness of our approach in efficiently managing memory during latent optimization, while previous methods fall short. Our proposed approach not only minimizes memory requirements but also significantly enhances T2V performance across a range of diffusion guidance benchmarks. Code, models, and demo are available at this https URL.
Abstract (translated)
近期，在条件扩散模型的发展中，为了在某一类任务上进行控制，仍然需要大量的监督微调。通过现成模型的引导来进行无需训练的条件设置是一种有利的选择，可以避免对基础模型进一步进行微调。然而，现有的无需训练的引导框架要么有较大的内存需求，要么由于粗略估计而导致提供次优的控制效果。这些缺点限制了其在需要大量计算的任务控制扩散模型（如文本到视频(Text-to-Video, T2V)扩散模型）中的应用。 为此，我们提出了一种名为TITAN-Guide的方法，用于引导式文本到视频扩散模型的时间推理对齐技术，该方法解决了内存空间问题，并且相比现有方法，在指导过程中提供了更优的控制效果。特别是，我们开发出一种无需通过判别性引导模型进行反向传播来优化扩散潜在变量的有效方法。特别地，我们在各种方向指令选项下研究了引导式扩散任务中的正向梯度下降法。 在实验中，我们展示了我们的方法在处理潜在变量优化时有效管理内存的优势，而先前的方法在这方面表现不足。我们提出的方法不仅将内存需求最小化，而且显著提高了多种扩散指导基准测试下的T2V性能。代码、模型和演示可在以下链接获得：[此URL]。 请注意，在原始文本中有一个未提供的URL，建议您访问该链接以获取更多的信息和技术细节。
URL
https://arxiv.org/abs/2508.00289
PDF
-
Whisker-based Active Tactile Perception for Contour Reconstruction
2025-07-31 07:38:42Yixuan Dang, Qinyang Xu, Yu Zhang, Xiangtong Yao, Liding Zhang, Zhenshan Bing, Florian Roehrbein, Alois KnollAbstract
Perception using whisker-inspired tactile sensors currently faces a major challenge: the lack of active control in robots based on direct contact information from the whisker. To accurately reconstruct object contours, it is crucial for the whisker sensor to continuously follow and maintain an appropriate relative touch pose on the surface. This is especially important for localization based on tip contact, which has a low tolerance for sharp surfaces and must avoid slipping into tangential contact. In this paper, we first construct a magnetically transduced whisker sensor featuring a compact and robust suspension system composed of three flexible spiral arms. We develop a method that leverages a characterized whisker deflection profile to directly extract the tip contact position using gradient descent, with a Bayesian filter applied to reduce fluctuations. We then propose an active motion control policy to maintain the optimal relative pose of the whisker sensor against the object surface. A B-Spline curve is employed to predict the local surface curvature and determine the sensor orientation. Results demonstrate that our algorithm can effectively track objects and reconstruct contours with sub-millimeter accuracy. Finally, we validate the method in simulations and real-world experiments where a robot arm drives the whisker sensor to follow the surfaces of three different objects.
Abstract (translated)
当前，基于触须灵感的触觉传感器在感知领域面临的主要挑战是机器人缺乏对直接接触信息进行主动控制的能力。为了精确地重建物体轮廓，触须传感器需要持续跟踪并保持适当的相对接触姿态。这尤其对于依赖于尖端接触定位的方法来说至关重要，因为这类方法对于锋利表面具有较低的容忍度，并且必须避免进入切向接触状态。 在本文中，我们首先构建了一种磁感应式触须传感器，该传感器配备了由三个灵活螺旋臂组成的紧凑而坚固的悬挂系统。接下来，开发了一种利用已知触须偏转特征直接通过梯度下降法提取尖端接触位置的方法，并应用贝叶斯滤波来减少波动。此外，我们还提出了一种主动运动控制策略，以保持触须传感器相对于物体表面的最佳相对姿态。 为了预测局部曲率并确定传感器的方向，使用了B样条曲线。实验结果表明，我们的算法能够有效地跟踪对象并实现次毫米精度的轮廓重建。最后，在模拟和真实世界的试验中验证了该方法的有效性——在这些试验中，机器人手臂驱动触须传感器沿三个不同物体表面进行接触跟踪。 总的来说，这项研究为利用机械式触觉传感器执行复杂的定位任务提供了一种新的途径，并有助于改善基于直接接触的感知系统。
URL
https://arxiv.org/abs/2507.23305
PDF
-
Efficient Differentially Private Fine-Tuning of LLMs via Reinforcement Learning
2025-07-30 10:46:53Afshin Khadangi, Amir Sartipi, Igor Tchappi, Ramin Bahmani, Gilbert FridgenAbstract
The tension between data privacy and model utility has become the defining bottleneck for the practical deployment of large language models (LLMs) trained on sensitive corpora including healthcare. Differentially private stochastic gradient descent (DP-SGD) guarantees formal privacy, yet it does so at a pronounced cost: gradients are forcibly clipped and perturbed with noise, degrading sample efficiency and final accuracy. Numerous variants have been proposed to soften this trade-off, but they all share a handicap: their control knobs are hard-coded, global, and oblivious to the evolving optimization landscape. Consequently, practitioners are forced either to over-spend privacy budget in pursuit of utility, or to accept mediocre models in order to stay within privacy constraints. We present RLDP, the first framework to cast DP optimization itself as a closed-loop control problem amenable to modern deep reinforcement learning (RL). RLDP continuously senses rich statistics of the learning dynamics and acts by selecting fine-grained per parameter gradient-clipping thresholds as well as the magnitude of injected Gaussian noise. A soft actor-critic (SAC) hyper-policy is trained online during language model fine-tuning; it learns, from scratch, how to allocate the privacy budget where it matters and when it matters. Across more than 1,600 ablation experiments on GPT2-small, Llama-1B, Llama-3B, and Mistral-7B, RLDP delivers perplexity reductions of 1.3-30.5% (mean 5.4%) and an average 5.6% downstream utility gain. RLDP reaches each baseline's final utility after only 13-43% of the gradient-update budget (mean speed-up 71%), all while honoring the same ($\epsilon$, $\delta$)-DP contract and exhibiting equal or lower susceptibility to membership-inference and canary-extraction attacks.
Abstract (translated)
在敏感语料库（包括医疗保健）上训练的大规模语言模型（LLMs）的实际部署中，数据隐私与模型效用之间的紧张关系已成为主要瓶颈。差分隐私随机梯度下降（DP-SGD）提供了正式的隐私保证，但代价是显著的：它强制将梯度裁剪并添加噪声，从而降低样本效率和最终准确性。尽管提出了许多变体以缓解这种权衡，但它们都存在一个共同的问题：其控制参数是硬编码、全局性的，并且无视不断变化的优化景观。因此，实践者要么被迫过度使用隐私预算来追求效用，要么接受中等水平的质量模型以符合隐私限制。 我们介绍了RLDP框架，这是第一个将差分隐私优化本身作为现代深度强化学习（RL）可处理的闭环控制问题进行建模的方法。RLDP通过选择细化的逐参数梯度裁剪阈值以及注入高斯噪声的数量来连续感知学习动态的丰富统计信息并采取行动。在语言模型微调期间在线训练软演员评论家（SAC）超策略；它从零开始学习如何在其重要的领域和重要时刻分配隐私预算。 在GPT2-small、Llama-1B、Llama-3B 和 Mistral-7B 上进行的超过 1,600 次消融实验中，RLDP 实现了困惑度降低了 1.3% 到 30.5%（平均降低 5.4%），并且下游效用平均提高了 5.6%。RLDP 在仅使用每个基线最终效用所需的梯度更新预算的13%-43%时就达到了每个基准的最终效用，同时遵守相同的($\epsilon$, $\delta$)-差分隐私合同，并且表现出相同或更低的成员推断和金丝雀提取攻击易受性。
URL
https://arxiv.org/abs/2507.22565
PDF
