Depth-wise pruning accelerates LLM inference in resource-constrained scenarios but suffers from performance degradation due to indiscriminate removal of entire Transformer layers. This paper reveals ``Patch-Like'' redundancy across layers via correlation analysis of the outputs of different layers in reproducing kernel Hilbert space, demonstrating consecutive layers exhibit high functional similarity. Building on this observation, this paper proposes Sliding-Window Merging (SWM) - a dynamic compression method that selects consecutive layers from top to bottom using a pre-defined similarity threshold, and compacts patch-redundant layers through a parameter consolidation, thereby simplifying the model structure while maintaining its performance. Extensive experiments on LLMs with various architectures and different parameter scales show that our method outperforms existing pruning techniques in both zero-shot inference performance and retraining recovery quality after pruning. In particular, in the experiment with 35\% pruning on the Vicuna-7B model, our method achieved a 1.654\% improvement in average performance on zero-shot tasks compared to the existing method. Moreover, we further reveal the potential of combining depth pruning with width pruning to enhance the pruning effect.
论文检索
输入标题、作者或关键词，从 1,235 篇学术成果中精准定位
Enhancing Kernel Power $K$-means: Scalable and Robust Clustering with Random Fourier Features and Possibilistic Method
PDF ↗Kernel power k-means (KPKM) leverages a family of means to mitigate local minima issues in kernel k-means. However, KPKM faces two key limitations: (1) the computational burden of the full kernel matrix restricts its use on extensive data, and (2) the lack of authentic centroid-sample assignment learning reduces its noise robustness. To overcome these challenges, we propose RFF-KPKM, introducing the first approximation theory for applying random Fourier features (RFF) to KPKM. RFF-KPKM employs RFF to generate efficient, low dimensional feature maps, bypassing the need for the whole kernel matrix. Crucially, we are the first to establish strong theoretical guarantees for this combination: (1) an excess risk bound of O( k^3/n), (2) strong consistency with membership values, and (3) a (1 + ε) relative error bound achievable using the RFF of dimension poly(ε^{−1} logk). Furthermore, to improve robustness and the ability to learn multiple kernels, we propose IP-RFF-MKPKM, an improved possibilistic RFF-based multiple kernel power k-means. IP-RFF-MKPKM ensures the scalability of MKPKM via RFF and refines cluster assignments by combining the merits of the possibilistic and fuzzy membership. Experiments on large-scale datasets demonstrate the superior efficiency and clustering accuracy of the proposed methods compared to the state-of-the-art alternatives.
Detecting unobserved confounders is crucial for reliable causal inference in observational studies. Existing methods require either linearity assumptions or multiple heterogeneous environments, limiting applicability to nonlinear single-environment settings. To bridge this gap, we propose Kernel Regression Confounder Detection (KRCD), a novel method for detecting unobserved confounding in nonlinear observational data under single-environment conditions. KRCD leverages reproducing kernel Hilbert spaces to model complex dependencies. By comparing standard and higher-order kernel regressions, we derive a test statistic whose significant deviation from zero indicates unobserved confounding. Theoretically, we prove two key results: First, in infinite samples, regression coefficients coincide if and only if no unobserved confounders exist. Second, finite-sample differences converge to zero-mean Gaussian distributions with tractable variance. Extensive experiments on synthetic benchmarks and the Twins dataset demonstrate that KRCD not only outperforms existing baselines but also achieves superior computational efficiency.
At present, spectral clustering is an important branch of unsupervised learning, and its application in deep learning has been widely concerned. However, for high-dimensional sparse datasets, the complexity of network scale leads to parameter explosion, and static Gaussian kernel often has wrong preset data structure. To overcome these challenges, we propose a novel deep clustering model, Deep Clustering Based on Sparse Kolmogorov-Arnold Network (KAN) and Spectral Constraint. It contains a deep sparse clustering framework, in which sparse KAN and the orthogonal layer are designed to enhance the sparsity of the activation function matrix, reduce the number of parameters and improve the stability of model convergence. Additionally, we add an adaptive optimized affinity matrix based on spectral constraint, which overcomes the limitations of static Gaussian kernels, and improves the performance and stability of spectral constraint. Experimental results on both synthetic and real datasets demonstrate that our model outperforms existing methods in clustering performance, computational efficiency, and stability.
Two Heads Are Better than One: Distilling Large Language Model Features into Small Models with Feature Decomposition and Mixture
PDF ↗Market making (MM) through Reinforcement Learning (RL) has attracted significant attention in financial trading. With the development of Large Language Models (LLMs), more and more attempts are being made to apply LLMs to financial areas. A simple, direct application of LLM as an agent shows significant performance. Such methods are hindered by their slow inference speed, while most of the current research has not studied LLM distillation for this specific task. To address this, we first propose the normalized fluorescent probe to study the mechanism of the LLM’s feature. Based on the observation found by our investigation, we propose Cooperative Market Making (CMM), a novel framework that decouples LLM features across three orthogonal dimensions: layer, task, and data. Various student models collaboratively learn simple LLM features along with different dimensions, with each model responsible for a distinct feature to achieve knowledge distillation. Furthermore, CMM introduces an Hájek-MoE to integrate the output of the student models by investigating the contribution of different models in a kernel function-generated common feature space. Extensive experimental results on four real-world market datasets demonstrate the superiority of CMM over the current distillation method and RL-based market-making strategies.
In large-scale sensor networks, Multivariate Time Series Classification (MTSC) is a pivotal task for identifying events dependent on longitudinal data at the edge. However, existing methods focus on neither the inherent ability of convolutional networks to perceive subsequence features, nor the prolonged processing pipeline and the model deployment overhead brought by the highly parameterized models. To resolve these difficulties, we present EdgeMTSC, a lightweight large-kernel ConvNet for MTSC, which naturally extracts and learns features of diverse subsequences. Specifically, a novel module named Inter-channel Message Passing-driven Kernel Block (IMP-KB) is proposed, which maintains a learnable correlation matrix to propagate and merge inter-channel messages, and fuses miscellaneous patterns learned by parallel conv kernels of different sizes. EdgeMTSC sequences two modules of different receptive fields to aggregate local features using small kernels and study long-term representation provided by large kernels, respectively. For inference parameter reduction and accelerating inference without performance loss, the conv blocks in IMP-KBs are structurally reparameterizable. The performance of our model (76.2%) is benchmarked on 26 UEA MTSC datasets and is superior to the SOTA model (MPTSNet, 75%). At the same time, EdgeMTSC uses the fewest parameters and achieves the minimum inference time, applicable on any machine (8 devices ranging from large-scale distributed AI computing servers to resource-constrained edge devices) and in any application scenario.
Anomaly detection on data streams presents significant challenges, requiring methods to maintain high detection accuracy among evolving distributions while ensuring real-time efficiency. Here we introduce IDK-S, a novel Incremental Distributional Kernel for Streaming anomaly detection that effectively addresses these challenges by creating a new dynamic representation in the kernel mean embedding framework. The superiority of IDK-S is attributed to two key innovations. First, it inherits the strengths of the Isolation Distributional Kernel, an offline detector that has demonstrated significant performance advantages over foundational methods like Isolation Forest and Local Outlier Factor due to the use of a data-dependent kernel. Second, it adopts a lightweight incremental update mechanism that significantly reduces computational overhead compared to the naive baseline strategy of performing a full model retraining. This is achieved without compromising detection accuracy, a claim supported by its statistical equivalence to the full retrained model. Our extensive experiments on thirteen benchmarks demonstrate that IDK-S achieves superior detection accuracy while operating substantially faster, in many cases by an order of magnitude, than existing state-of-the-art methods.
Sat2Flow: A Structure-Aware Diffusion Framework for Human Flow Generation from Satellite Imagery
PDF ↗Origin-Destination (OD) flow matrices are critical for urban mobility analysis, supporting traffic forecasting, infrastructure planning, and policy design. Existing methods face two key limitations: (1) reliance on costly auxiliary features (e.g., Points of Interest, socioeconomic statistics) with limited spatial coverage, and (2) fragility to spatial topology changes, where reordering urban regions disrupts the structural coherence of generated flows. We propose Sat2Flow, a structure-aware diffusion framework that generates structurally coherent OD flows using only satellite imagery. Our approach employs a multi-kernel encoder to capture diverse regional interactions and a permutation-aware diffusion process that maintains consistency across regional orderings. Through joint contrastive training linking satellite features with OD patterns and equivariant diffusion training enforcing structural invariance, Sat2Flow ensures topological robustness under arbitrary regional reindexing. Experiments on real-world datasets show that Sat2Flow outperforms physics-based and data-driven baselines in accuracy while preserving flow distributions and spatial structures under index permutations. Sat2Flow offers a globally scalable solution for OD flow generation in data-scarce environments, eliminating region-specific auxiliary data dependencies while maintaining structural robustness for reliable mobility modeling.
Despite substantial progress in anomaly synthesis, existing diffusion-based and coarse inpainting pipelines commonly suffer from structural deficiencies such as micro-structural discontinuities, limited semantic controllability, and inefficient generation. To overcome these limitations, we introduce ARAS, a language-conditioned, auto-regressive anomaly synthesis approach that precisely injects local, text-specified defects into normal images via token-anchored latent editing. Leveraging a hard-gated auto-regressive operator and a training-free, context-preserving masked sampling kernel, ARAS significantly enhances defect realism, preserves fine-grained material textures, and provides continuous semantic control over synthesized anomalies. Integrated within our Quality-Aware Re-weighted Anomaly Detection (QARAD) framework, we propose a dynamic weighting strategy that emphasizes high-quality synthetic samples by computing an image-text similarity score with a dual-encoder model. Extensive experiments across three datasets, MVTec AD, VisA, and BTAD, demonstrate that our QARAD outperforms SOTA methods in both image- and pixel-level anomaly detection tasks, achieving improved accuracy, robustness, and a 5× synthesis speedup compared to diffusion-based alternatives.
Video recommendation systems heavily rely on user watch time feedback, making accurate watch time prediction a crucial task. However, this task inherently suffers from bias, as recommendation models tend to favor long-duration videos to maximize watch time. This issue, known as duration bias in the watch-time prediction context, can be explained from a causal perspective, where video duration acts as a confounder. Recent works address this bias using backdoor adjustment, isolating the direct effect of content on watch time from observational data. These methods typically discretize video duration into groups, estimate group-wise effects, and then aggregate them via a unified prediction model. However, this aggregation strategy is prone to model misspecification due to feature distribution shift across groups. In this paper, we reinterpret the problem through the lens of invariant learning and propose a novel framework: Duration-Invariant Feature Learning (DIFL). DIFL employs a kernel-based regularization that enforces representation invariance across duration groups, reducing sensitivity to group design and improving generalization. This enables more accurate modeling of the direct causal effect and making counterfactual inference. Extensive experiments on both public and real large-scale production datasets demonstrate the effectiveness of our approach, which achieves SOTA performance.
Graph Contrastive Learning (GCL) has emerged as a powerful paradigm for training Graph Neural Networks (GNNs) in the absence of task-specific labels. However, its scalability on large-scale graphs is hindered by the intensive message passing mechanism of GNN and the quadratic computational complexity of contrastive loss over positive and negative node pairs. To address these issues, we propose an efficient GCL framework that transforms the input graph into a compact network of interconnected node sets while preserving structural information across communities. We firstly introduce a kernelized graph community contrastive loss with linear complexity, enabling effective information transfer among node sets to capture hierarchical structural information of the graph. We then incorporate a knowledge distillation technique into the decoupled GNN architecture to accelerate inference while maintaining strong generalization performance. Extensive experiments on sixteen real-world datasets of varying scales demonstrate that our method outperforms state-of-the-art GCL baselines in both effectiveness and scalability.
There is a growing demand for deploying large generative AI models on mobile devices. For recent popular video generative models, however, the Variational AutoEncoder (VAE) represents one of the major computational bottlenecks. Both large parameter sizes and mismatched kernels cause out-of-memory errors or extremely slow inference on mobile devices. To address this, we propose a low-cost solution that efficiently transfers widely used video VAEs to mobile devices. (1) We analyze redundancy in existing VAE architectures and get empirical design insights. By integrating 3D depthwise separable convolutions into our model, we significantly reduce the number of parameters. (2) We observe that the upsampling techniques in mainstream video VAEs are poorly suited to mobile hardware and form the main bottleneck. In response, we propose a decoupled 3D pixel shuffle scheme that slashes end-to-end delay. Building upon these, we develop a universal mobile-oriented VAE decoder, Turbo-VAED. (3) We propose an efficient VAE decoder training method. Since only the decoder is used during deployment, we distill it to Turbo-VAED instead of retraining the full VAE, enabling fast mobile adaptation with minimal performance loss. To our knowledge, our method enables real-time 720p video VAE decoding on mobile devices for the first time. This approach is widely applicable to most video VAEs. When integrated into four representative models, with training cost as low as $95, it accelerates original VAEs by up to 84.5× at 720p resolution on GPUs, uses as low as 17.5% of original parameter count, and retains 96.9% of the original reconstruction quality. Compared to mobile-optimized VAEs, Turbo-VAED achieves a 2.9× speedup in FPS and better reconstruction quality on the iPhone 16 Pro.
In this paper, we show that current approaches using large square kernels or transformer-based global modeling aggregate contextual information uniformly across spatial dimensions, leading to feature dilution and localization errors for elongated targets. To mitigate this issue, we propose Strip R-CNN, the first work to systematically explore large strip convolutions for remote sensing object detection. Our key insight is that strip convolutions enable directional feature aggregation along the dominant spatial dimension of slender objects, reducing background interference while preserving essential geometric information. We design two core components: (i) StripNet, a backbone network employing sequential orthogonal large strip convolutions to capture anisotropic spatial patterns, and (ii) Strip Head, which enhances localization precision by incorporating strip convolutions into the detection head. Unlike previous large-kernel approaches that suffer from computational redundancy and isotropic limitations, our method achieves superior performance with remarkable efficiency. Extensive experiments on multiple benchmarks (DOTA, FAIR1M, HRSC2016, and DIOR) demonstrate significant improvements, with our 30M parameter model achieving 82.75% mAP on DOTA-v1.0, establishing a new state-of-the-art record while providing new insights into anisotropic feature learning for remote sensing applications.
Conformable Convolution for Topologically Constrained Learning of Complex Anatomical Structures
PDF ↗While conventional computer vision emphasizes pixel-level and feature-based objectives, medical image analysis of intricate biological structures necessitates explicit representation of their complex topological properties. Despite their successes, deep learning models often struggle to accurately capture the connectivity and continuity of fine, sometimes pixel-thin, yet critical structures due to their reliance on implicit learning from data. To address this challenge, we introduce Conformable Convolution, a novel convolutional layer designed to explicitly impose topological consistency. Conformable Convolution learns adaptive kernel offsets that focus on regions of high topological significance within an image. This prioritization is guided by our proposed Topological Posterior Generator (TPG) module, which leverages persistent homology. The TPG module identifies key topological features and guides the convolutional layers by applying persistent homology to feature maps transformed into cubical complexes. Unlike existing approaches that are merely aware of topology, our method explicitly constrains the learning process to ensure topological correctness. The proposed modules are architecture-agnostic, enabling them to be integrated seamlessly into various architectures. We showcase the effectiveness of our framework in the segmentation task, where preserving the interconnectedness of structures is critical. The results on three diverse datasets demonstrate that our framework effectively preserves the topology both quantitatively and qualitatively.
Recently, Test-Time Adaptation (TTA) has gained increasing attention in medical imaging due to its ability to improve model generalization under domain shifts without retraining. In particular, directly applying a well-trained model across various medical centers faces significant performance degradation caused by variations in equipment, operators, imaging conditions, and scanning skill levels of sonographers. Existing TTA methods either rely on parameter adaptation that increases computational cost or apply simple prediction fusion that ignores anatomical structure knowledge. To address these limitations, we propose a novel backward-free Topology-aware TTA framework named T^3 that integrates Structural Perception Modeling (SPM) and Box Regression Adaptation (BRA). SPM is implemented through an organ space heatmap generated via Gaussian kernel superposition. This heatmap encodes anatomical topology without requiring additional training or source data. BRA further improves localization and classification by fusing detection outputs based on the contribution of detected results to anatomically meaningful peak points from the heatmaps. Extensive experiments were conducted across six cross-domain scenarios, and the results demonstrate that our method achieves state-of-the-art cross-domain detection performance while maintaining high efficiency, offering a practical and robust solution for real-world medical diagnostic applications.
Ground What You See: Hallucination-Resistant MLLMs via Caption Feedback, Diversity-Aware Sampling, and Conflict Regularization
PDF ↗Multimodal large language models (MLLMs) have achieved significant results in various tasks, but their practical application is still severely constrained by hallucination issues, which are particularly prominent in reinforcement learning (RL) optimization processes. This paper systematically analyzes the causes of hallucinations in MLLM under RL training, identifying three key factors: (1) The model relies heavily on chained visual reasoning to guide decision-making during RL training. Thus, error and irrelevant information in visual reasoning can easily cause hallucinations, including inaccurate initial visual descriptions that anchor subsequent inferences to incorrect information, as well as redundant and broad inferential information; (2) Insufficient exploration diversity during the policy optimization phase, causing the model to output overly confident results; (3) The destructive conflict between different samples during optimization is a key factor that leads to false associations and unstable parameter updates. To address these issues, we propose a solution framework comprising three core modules. First, to improve the accuracy of visual localization, we add planning and caption stages before thinking and answer stages. To enhance initial visual descriptions ability, we allow LLMs to respond based solely on the caption and provide corresponding caption reward based on the quality of the response. Second, to enhance exploration capabilities, we classify samples based on the mean and variance of the reward distribution and select samples with high reward variance for training, thereby increasing the model's focus on diverse samples. Finally, to mitigate conflicts between training samples, we identify neural tangent kernel (NTK) similarity as the key factor. Rather than minimizing it uniformly, we regulate NTK similarity by grouping sample pairs based on a similarity threshold. An InfoNCE loss is then applied to pull dissimilar pairs closer and push overly similar ones apart, guiding interactions toward a balanced range. The experimental results demonstrate that the proposed method significantly reduces the hallucination rate and effectively improves the inference accuracy of MLLMs.
AquaSplatting: A Hybrid 3D Representation for Robust Underwater Scene Reconstruction via Dual-Branch Rendering
PDF ↗While 3D Gaussian Splatting (3DGS) excels at real-time rendering of standard scenes, it struggles to reconstruct underwater environments due to severe challenges such as light scattering, color attenuation, and sparse coverage of Gaussian kernels in far-field aqueous regions. To address this, we introduce AquaSplatting, a hybrid framework that combines explicit and implicit modeling methods for robust underwater scene reconstruction. Our dual-branch architecture employs 3DGS in a geometry-guided branch to model solid surfaces like the seabed, while a medium-aware branch uses a compact, view-dependent MLP to represent volumetric water effects. Furthermore, a neural underwater hybrid rendering mechanism adaptively fuses these two representations based on accumulated opacity. Thanks to this dual-branch framework, our method can also synthesize restored images without water medium. To enhance efficiency, our proposed engagement-based pruning (EBP) strategy quantifies each Gaussian's contribution by accumulating its image-space gradients over multiple frames, enabling the principled removal of primitives with negligible impact. The entire framework is optimized using a comprehensive loss function that integrates photometric, exposure, semantic, and depth priors to maximize visual fidelity. Experiments on challenging underwater datasets demonstrate that AquaSplatting achieves the state-of-the-art in reconstruction quality surpassing prior methods while maintaining real-time performance.
Visible-Infrared (RGB-IR) Unmanned Aerial Vehicle (UAV) object detection integrates complementary cues from visible and infrared sensors, offering broad application potential. However, due to sensor parallax, it still faces the challenge of weak spatial misalignment, which significantly limits its performance in UAV-based object detection. Existing methods emphasize strict alignment, overlooking spectral heterogeneity under varying illumination. To address these issues, we propose the Illumination Guided Implicit Alignment Network (IGIANet) to mitigate modality heterogeneity without explicit alignment. Specifically, we integrate three novel modules. First, we propose an illumination-guided frequency modulation module that adaptively allocates fusion weights to visible and infrared features based on global illumination estimation, effectively alleviating modality imbalance under varying lighting conditions. Second, we introduce a frequency-guided cross-modality differential enhancement module, which computes differential cues across frequency domains to enhance complementary information and highlight weakly aligned and low-contrast regions. Finally, we introduce an implicit alignment-driven dynamic fusion module that actively estimates offsets and generates dynamic, position-adaptive fusion kernels to align and fuse modalities. Extensive experiments demonstrate that IGIANet outperforms state-of-the-art models on various benchmarks, achieving 80.9% mAP on DroneVehicle, 57.1% mAP on VEDAI, and 49.4% mAP on FLIR.
Monte Carlo random walk methods are widely used in capacitance extraction for their mesh-free formulation and inherent parallelism. However, modern semiconductor technologies with densely packed structures present significant challenges in unbiasedly sampling transition domains in walk steps with multiple high-contrast dielectric materials. We present DeepRWCap, a machine learning-guided random walk solver that predicts the transition quantities required to guide each step of the walk. These include Poisson kernels, gradient kernels, signs and magnitudes of weight. DeepRWCap employs a two-stage neural architecture that decomposes structured outputs into face-wise distributions and spatial kernels on cube faces. It uses 3D convolutional networks to capture volumetric dielectric interactions and 2D depthwise separable convolutions to model localized kernel behavior. The design incorporates grid-based positional encodings and structural design choices informed by cube symmetries to reduce learning redundancy and improve generalization. Trained on 100,000 procedurally generated dielectric configurations, DeepRWCap achieves a mean relative error of 1.24±0.53% when benchmarked against the commercial Raphael solver on the self-capacitance estimation of 10 industrial designs spanning 12 to 55 nm nodes. Compared to the state-of-the-art stochastic difference method Microwalk, DeepRWCap achieves an average 23% speedup. On complex designs with runtimes over 10s, it reaches an average 49% acceleration.
Neural operators have emerged as a promising paradigm for learning solution operators of partial differential equations (PDEs) directly from data. Existing methods, such as those based on Fourier or graph techniques, make strong assumptions about the structure of the kernel integral operator, assumptions which may limit expressivity. We present SVD-NO, a neural operator that explicitly parameterizes the kernel by its singular-value decomposition (SVD) and then carries out the integral directly in the low-rank basis. Two lightweight networks learn the left and right singular functions, a diagonal parameter matrix learns the singular values, and a Gram-matrix regularizer enforces orthonormality. As SVD-NO approximates the full kernel, it obtains a high degree of expressivity. Furthermore, due to its low-rank structure the computational complexity of applying the operator remains reasonable, leading to a practical system. In extensive evaluations on five diverse benchmark equations, SVD-NO achieves a new state of the art. In particular, SVD-NO provides greater performance gains on PDEs whose solutions are highly spatially variable.