arXiv subjects

Shiwei Wang

Publications and source records attributed to Shiwei Wang.

At least 19 recordsLinked to original sources

arXiv2026

Detecting Backdoors in Object Detection via Pre-NMS Prediction Distribution Shift

Object detection models deployed in safety-critical applications remain vulnerable to backdoor attacks that cause targeted misbehaviors when a hidden trigger is present. Existing detection methods either rely on trigger inversion or exploit architecture-specific assumptions, and critically, representative existing methods fail to generalize reliably to scene-level attacks, where a single trigger induces anomalous behavior across all objects in the scene simultaneously. We present DistScan, a backdoor detection framework based on a simple but previously unexploited observation: backdoor injection systematically shifts a model's pre-NMS prediction class distribution away from its training class frequencies, even on clean inputs without any trigger present. DistScan aggregates intermediate class predictions over a clean validation set and flags a model as backdoored if the resulting distribution deviates significantly from the training class frequencies, requiring no model weight access, no trigger knowledge, and no additional training. Extensive experiments on MS-COCO and PASCAL VOC across two architectures and three scene-level attack scenarios demonstrate that DistScan substantially outperforms existing methods, improving average detection accuracy over the best-performing applicable baseline by 27.32 percentage points.

cs.CV↗

arXiv2026

Hydrogen-induced lattice cohesion weakening favors atomic displacement

Atomic displacement -- the fundamental process underlying diverse deformation and damage phenomena in metals, from irradiation defect production to stress-driven dislocation motion -- is governed by interatomic cohesion strength. Here, lattice-dissolved hydrogen (LDH) occurring in metals under direct hydrogen exposure is identified to effectively weaken lattice cohesion, and thereby facilitating atomic displacement and dislocation movement upon plastic deformation in sub-threshold stress regime. This atomic-scale insight provides a physically transparent mechanism for hydrogen-enhanced localized plasticity implicated in hydrogen embrittlement. We quantitatively verify the hydrogen-induced lattice cohesion weakening effect on metal surfaces exposed to low-energy hydrogen plasma, where massive defects are generated despite the absence of sufficient ion momentum for direct displacement damage. By unprecedentedly quantifying the cohesion-weakening effect of LDH independently from defect-trapped H, we establish a new paradigm to understand hydrogen embrittlement.

cond-mat.mtrl-sci↗

arXiv2026

ASTRA-QA: A Benchmark for Abstract Question Answering over Documents

Document-based question answering (QA) increasingly includes abstract questions that require synthesizing scattered information from long documents or across multiple documents into coherent answers. However, this setting is still poorly supported by existing benchmarks and evaluation methods, which often lack stable abstract references or rely on coarse similarity metrics and unstable head-to-head comparisons. To alleviate this issue, we introduce ASTRA-QA, a benchmark for AbSTRAct Question Answering over documents. ASTRA-QA contains 869 QA instances over academic papers and news documents, covering five abstract question types and three controlled retrieval scopes. Each instance is equipped with explicit evaluation annotations, including answer topic sets, curated unsupported topics, and aligned evidence. Building on these annotations, ASTRA-QA assesses whether answers cover required key points and avoid unsupported content by directly scoring topic coverage and curated unsupported content, enabling scalable evaluation without exhaustive head-to-head comparisons. Experiments with representative Retrieval-Augmented Generation (RAG) methods spanning vanilla, graph-based, and hierarchical retrieval settings show that ASTRA-QA provides reference-grounded diagnostics for coverage, hallucination, and retrieval-scope robustness. Our dataset and code are available at https://xinyangsally.github.io/astra-benchmark.

cs.CL↗

arXiv2026

M2IR: Proactive All-in-One Image Restoration via Mamba-style Modulation and Mixture-of-Experts

While Transformer-based architectures have dominated recent advances in all-in-one image restoration, they remain fundamentally reactive: propagating degradations rather than proactively suppressing them. In the absence of explicit suppression mechanisms, degraded signals interfere with feature learning, compelling the decoder to balance artifact removal and detail preservation, thereby increasing model complexity and limiting adaptability. To address these challenges, we propose M2IR, a novel restoration framework that proactively regulates degradation propagation during the encoding stage and efficiently eliminates residual degradations during decoding. Specifically, the Mamba-Style Transformer (MST) block performs pixel-wise selective state modulation to mitigate degradations while preserving structural integrity. In parallel, the Adaptive Degradation Expert Collaboration (ADEC) module utilizes degradation-specific experts guided by a DA-CLIP-driven router and complemented by a shared expert to eliminate residual degradations through targeted and cooperative restoration. By integrating the MST block and ADEC module, M2IR transitions from passive reaction to active degradation control, effectively harnessing learned representations to achieve superior generalization, enhanced adaptability, and refined recovery of fine-grained details across diverse all-in-one image restoration benchmarks. Our source codes are available at https://github.com/Im34v/M2IR.

cs.CV↗

arXiv2026

On the efficient computation of proximal operators of affine-constrained nonconvex functions

Proximal operators with affine constraints arise in numerous models in nonconvex projection, composite optimization, and structured regularization. However, their efficient computation remains challenging due to the simultaneous presence of affine constraints and nonsmooth, possibly nonconvex objectives. In this work, we develop a unified dual-representability framework for analyzing and computing affine-constrained proximal mappings. Specifically, we introduce a multiplier inclusion formulation that connects the primal affine-constrained proximal problem to an unconstrained convex dual problem. Based on this formulation, we prove that, whenever the associated dual inclusion problem admits a solution, strong duality holds. For convex functions and a broad class of prox-regular nonconvex functions, we establish that dual representability holds under a simple subdifferential sum rule, and further develop a hierarchy of verifiable regularity conditions that guarantee this sum rule. In addition, we analyze the smoothness and strong convexity properties of the dual objective, providing a rigorous foundation that guarantees fast local convergence rates for efficient first- and second-order methods. Numerical experiments demonstrate that the proposed dual reformulation enables the reliable computation of globally optimal solutions for a range of large-scale nonconvex proximal and projection problems using existing convex optimization solvers.

math.OC↗

arXiv2026

Twice Epi-Differentiability of Spectral Functions and its applications

Second-order variational properties have been shown to play important theoretical and numerical roles for different classes of optimization problems. Among such properties, twice epi-differentiability has a special place because of its ubiquitous presence in various classes of extended-real-valued functions that are important for optimization problems. We provide a useful characterization of this property for spectral functions by demonstrating that it can be characterized via the same property of the symmetric part of the spectral representation of an eigenvalue function. Our approach allows us to bypass the rather restrictive convexity assumption, used in many recent works that targeted second-order variational properties of spectral functions. By this theoretical tool, several applications on the proto-differentiability of subgradient mappings, the directional differentiability of the proximal mapping of spectral functions are achieved. We finally use our established theory to study twice epi-differentiability of leading eigenvalue functions and practical regularization terms that have important applications in statistics and the robust PCA.

math.OC↗

arXiv2026

D-Legion: A Scalable Many-Core Architecture for Accelerating Matrix Multiplication in Quantized LLMs

The performance gains obtained by large language models (LLMs) are closely linked to their substantial computational and memory requirements. Quantized LLMs offer significant advantages with extremely quantized models, motivating the development of specialized architectures to accelerate their workloads. This paper proposes D-Legion, a novel scalable many-core architecture, designed using many adaptive-precision systolic array cores, to accelerate matrix multiplication in quantized LLMs. The proposed architecture consists of a set of Legions where each Legion has a group of adaptive-precision systolic arrays. D-Legion supports multiple computation modes, including quantized sparse and dense matrix multiplications. The block structured sparsity is exploited within a fully-sparse, or partially-sparse windows. In addition, memory accesses of partial summations (psums) are spatially reduced through parallel accumulators. Furthermore, data reuse is maximized through optimized scheduling techniques by multicasting matrix tiles across the Legions. A comprehensive design space exploration is performed in terms of Legion/core granularity to determine the optimal Legion configuration. Moreover, D-Legion is evaluated on attention workloads from two BitNet models, delivering up to 8.2$\times$ lower latency, up to 3.8$\times$ higher memory savings, and up to 3$\times$ higher psum memory savings compared to state-of-the-art work. D-Legion, with eight Legions and 64 total cores, achieves a peak throughput of 135.68 TOPS at a frequency of 1 GHz. A scaled version of D-Legion, with 32 Legions, is compared to Google TPUv4i, achieving up to 2.5$\times$ lower total latency, up to 2.3$\times$ higher total throughput, and up to 2.7$\times$ higher total memory savings.

cs.AR↗

arXiv2025

A Multi-Channel Auditory Signal Encoder with Adaptive Resolution Using Volatile Memristors

We demonstrate and experimentally validate an end-to-end hybrid CMOS-memristor auditory encoder that realises adaptive-threshold, asynchronous delta-modulation (ADM)-based spike encoding by exploiting the inherent volatility of HfTiOx devices. A spike-triggered programming pulse rapidly raises the ADM threshold Delta (desensitisation); the device's volatility then passively lowers Delta when activity subsides (resensitisation), emphasising onsets while restoring sensitivity without static control energy. Our prototype couples an 8-channel 130 nm encoder IC to off-chip HfTiOx devices via a switch interface and an off-chip controller that monitors spike activity and issues programming events. An on-chip current-mirror transimpedance amplifier (TIA) converts device current into symmetric thresholds, enabling both sensitive and conservative encoding regimes. Evaluated with gammatone-filtered speech, the adaptive loop-at matched spike budget-sharpens onsets and preserves fine temporal detail that a fixed-Delta baseline misses; multi-channel spike cochleagrams show the same trend. Together, these results establish a practical hybrid CMOS-memristor pathway to onset-salient, spike-efficient neuromorphic audio front-ends and motivate low-power single-chip integration.

eess.AS↗

arXiv2025

A Quadratically Convergent Alternating Projection Method for Nonconvex Sets

In this paper, we consider the feasibility problem, which aims to find a feasible point for the constraint set $\{x \in \mathbb{R}^n: c(x) = 0\}$ over a possibly non-regular subset $\mathcal{X} \subset \mathbb{R}^n$. Under the constraint nondegeneracy condition, we propose a modified alternating projection method. In our proposed method, based on the concept of projective mapping for $\mathcal{X}$, we alternate a Newton step for finding an inexact solution within the limiting tangent cone of $\mathcal{X}$ and a projection to $\mathcal{X}$. Under mild conditions, we prove the local quadratic convergence of our proposed method. Preliminary numerical experiments demonstrate the high efficiency of our proposed alternating projection method.

math.OC↗

arXiv2025

DISCA: A Digital In-memory Stochastic Computing Architecture Using A Compressed Bent-Pyramid Format

Nowadays, we are witnessing an Artificial Intelligence revolution that dominates the technology landscape in various application domains, such as healthcare, robotics, automotive, security, and defense. Massive-scale AI models, which mimic the human brain's functionality, typically feature millions and even billions of parameters through data-intensive matrix multiplication tasks. While conventional Von-Neumann architectures struggle with the memory wall and the end of Moore's Law, these AI applications are migrating rapidly towards the edge, such as in robotics and unmanned aerial vehicles for surveillance, thereby adding more constraints to the hardware budget of AI architectures at the edge. Although in-memory computing has been proposed as a promising solution for the memory wall, both analog and digital in-memory computing architectures suffer from substantial degradation of the proposed benefits due to various design limitations. We propose a new digital in-memory stochastic computing architecture, DISCA, utilizing a compressed version of the quasi-stochastic Bent-Pyramid data format. DISCA inherits the same computational simplicity of analog computing, while preserving the same scalability, productivity, and reliability of digital systems. Post-layout modeling results of DISCA show an energy efficiency of 3.59TOPS/W per bit at 500 MHz using a commercial 180 nm CMOS technology. Therefore, DISCA significantly improves the energy efficiency for matrix multiplication workloads by orders of magnitude if scaled and compared to its counterpart architectures.

cs.AR↗

arXiv2025

ADiP: Adaptive-Precision Systolic Array for Matrix Multiplication Acceleration

Transformers are at the core of modern AI nowadays. They rely heavily on matrix multiplication and require efficient acceleration due to their substantial memory and computational requirements. Quantization plays a vital role in reducing memory usage, and can be exploited for computations by designing reconfigurable architectures that enhance matrix multiplication by dynamically adjusting the precision. This paper proposes ADiP, a novel adaptive-precision systolic array architecture designed for efficient matrix multiplication acceleration. The proposed architecture consists of $N$ $\times$ $N$ reconfigurable processing elements (PEs), along with shared shifters and accumulators. ADiP supports multiple computation modes, including symmetric single-matrix multiplication as well as asymmetric multi-matrix multiplication with a shared input matrix, thereby improving data reuse and PE utilization. By adapting to different precisions, ADiP achieves up to 4$\times$ higher throughput and up to 4$\times$ higher memory efficiency. Analytical models are developed for ADiP architecture, including latency and throughput for different architecture configurations. A comprehensive hardware design space exploration is demonstrated using commercial 22nm technology. Furthermore, ADiP is evaluated on different Transformer-based workloads from GPT-2 medium, BERT large, and BitNet-1.58B models, delivering total latency improvement up to 53.6%, and total energy improvement up to 24.4% for attention workloads in BitNet-1.58B model. At a 64$\times$64 size with reconfigurable 4,096 PEs, ADiP achieves a peak throughput of 8.192 TOPS, 16.384 TOPS, and 32.768 TOPS for 8bit$\times$8bit, 8bit$\times$4bit, and 8bit$\times$2bit operations, respectively.

cs.AR↗

arXiv2025

QueryGenie: Making LLM-Based Database Querying Transparent and Controllable

Conversational user interfaces powered by large language models (LLMs) have significantly lowered the technical barriers to database querying. However, existing tools still encounter several challenges, such as misinterpretation of user intent, generation of hallucinated content, and the absence of effective mechanisms for human feedback-all of which undermine their reliability and practical utility. To address these issues and promote a more transparent and controllable querying experience, we proposed QueryGenie, an interactive system that enables users to monitor, understand, and guide the LLM-driven query generation process. Through incremental reasoning, real-time validation, and responsive interaction mechanisms, users can iteratively refine query logic and ensure alignment with their intent.

cs.HC↗

arXiv2025

An Exact Penalty Approach for Equality Constrained Optimization over a Convex Set

In this paper, we consider the nonlinear constrained optimization problem (NCP) with constraint set $\{x \in \mathcal{X}: c(x) = 0\}$, where $\mathcal{X}$ is a closed convex subset of $\mathbb{R}^n$. We propose an exact penalty approach, named constraint dissolving approach, that transforms (NCP) into its corresponding constraint dissolving problem (CDP). The transformed problem (CDP) admits $\mathcal{X}$ as its feasible region with a locally Lipschitz smooth objective function. We prove that (NCP) and (CDP) share the same first-order stationary points, second-order stationary points, second-order sufficient condition (SOSC) points, and strong SOSC points, in a neighborhood of the feasible region. Moreover, we prove that these equivalences extend globally under a particular error bound condition. Therefore, our proposed constraint dissolving approach enables direct implementations of optimization approaches over $\mathcal{X}$ and inherits their convergence properties to solve problems that take the form of (NCP). Preliminary numerical experiments illustrate the high efficiency of directly applying existing solvers for optimization over $\mathcal{X}$ to solve (NCP) through (CDP). These numerical results further demonstrate the practical potential of our proposed constraint dissolving approach.

math.OC↗

arXiv2025

Revisiting Training-Inference Trigger Intensity in Backdoor Attacks

Backdoor attacks typically place a specific trigger on certain training data, such that the model makes prediction errors on inputs with that trigger during inference. Despite the core role of the trigger, existing studies have commonly believed a perfect match between training-inference triggers is optimal. In this paper, for the first time, we systematically explore the training-inference trigger relation, particularly focusing on their mismatch, based on a Training-Inference Trigger Intensity Manipulation (TITIM) workflow. TITIM specifically investigates the training-inference trigger intensity, such as the size or the opacity of a trigger, and reveals new insights into trigger generalization and overfitting. These new insights challenge the above common belief by demonstrating that the training-inference trigger mismatch can facilitate attacks in two practical scenarios, posing more significant security threats than previously thought. First, when the inference trigger is fixed, using training triggers with mixed intensities leads to stronger attacks than using any single intensity. For example, on CIFAR-10 with ResNet-18, mixing training triggers with 1.0 and 0.1 opacities improves the worst-case attack success rate (ASR) (over different testing opacities) of the best single-opacity attack from 10.61\% to 92.77\%. Second, intentionally using certain mismatched training-inference triggers can improve the attack stealthiness, i.e., better bypassing defenses. For example, compared to the training/inference intensity of 1.0/1.0, using 1.0/0.7 decreases the area under the curve (AUC) of the Scale-Up defense from 0.96 to 0.62, while maintaining a high attack ASR (99.65\% vs. 91.62\%). The above new insights are validated to be generalizable across different backdoor attacks, models, datasets, tasks, and (digital/physical) domains.

cs.CR↗

arXiv2025

Characterizations of Tilt-Stable Local Minimizers of a Class of Matrix Optimization Problems

Tilt stability plays a pivotal role in understanding how local solutions of an optimization problem respond to small, targeted perturbations of the objective. Although quadratic bundles are a powerful tool for capturing second-order variational behavior, their characterization remains incomplete beyond well-known polyhedral and certain specialized nonpolyhedral settings. To help bridge this gap, we propose a new point-based criterion for tilt stability in prox-regular, subdifferentially continuous functions by exploiting the notion of minimal quadratic bundles. Furthermore, we derive an explicit formula for the minimal quadratic bundle associated with a broad class of general spectral functions, thus providing a practical and unifying framework that significantly extends existing results and offers broader applicability in matrix optimization problems.

math.OC↗

arXiv2025

High Dynamic Range Video Compression: A Large-Scale Benchmark Dataset and A Learned Bit-depth Scalable Compression Algorithm

Recently, learned video compression (LVC) is undergoing a period of rapid development. However, due to absence of large and high-quality high dynamic range (HDR) video training data, LVC on HDR video is still unexplored. In this paper, we are the first to collect a large-scale HDR video benchmark dataset, named HDRVD2K, featuring huge quantity, diverse scenes and multiple motion types. HDRVD2K fills gaps of video training data and facilitate the development of LVC on HDR videos. Based on HDRVD2K, we further propose the first learned bit-depth scalable video compression (LBSVC) network for HDR videos by effectively exploiting bit-depth redundancy between videos of multiple dynamic ranges. To achieve this, we first propose a compression-friendly bit-depth enhancement module (BEM) to effectively predict original HDR videos based on compressed tone-mapped low dynamic range (LDR) videos and dynamic range prior, instead of reducing redundancy only through spatio-temporal predictions. Our method greatly improves the reconstruction quality and compression performance on HDR videos. Extensive experiments demonstrate the effectiveness of HDRVD2K on learned HDR video compression and great compression performance of our proposed LBSVC network. Code and dataset will be released in https://github.com/sdkinda/HDR-Learned-Video-Coding.

cs.CV↗

arXiv2025

L-Sort: On-chip Spike Sorting with Efficient Median-of-Median Detection and Localization-based Clustering

Spike sorting is a critical process for decoding large-scale neural activity from extracellular recordings. The advancement of neural probes facilitates the recording of a high number of neurons with an increase in channel counts, arising a higher data volume and challenging the current on-chip spike sorters. This paper introduces L-Sort, a novel on-chip spike sorting solution featuring median-of-median spike detection and localization-based clustering. By combining the median-of-median approximation and the proposed incremental median calculation scheme, our detection module achieves a reduction in memory consumption. Moreover, the localization-based clustering utilizes geometric features instead of morphological features, thus eliminating the memory-consuming buffer for containing the spike waveform during feature extraction. Evaluation using Neuropixels datasets demonstrates that L-Sort achieves competitive sorting accuracy with reduced hardware resource consumption. Implementations on FPGA and ASIC (180 nm technology) demonstrate significant improvements in area and power efficiency compared to state-of-the-art designs while maintaining comparable accuracy. If normalized to 22 nm technology, our design can achieve roughly $\times 10$ area and power efficiency with similar accuracy, compared with the state-of-the-art design evaluated with the same dataset. Therefore, L-Sort is a promising solution for real-time, high-channel-count neural processing in implantable devices.

eess.SP↗

arXiv2024

E-Sort: Empowering End-to-end Neural Network for Multi-channel Spike Sorting with Transfer Learning and Fast Post-processing

Decoding extracellular recordings is a crucial task in electrophysiology and brain-computer interfaces. Spike sorting, which distinguishes spikes and their putative neurons from extracellular recordings, becomes computationally demanding with the increasing number of channels in modern neural probes. To address the intensive workload and complex neuron interactions, we propose E-Sort, an end-to-end neural network-based spike sorter with transfer learning and parallelizable post-processing. Our framework reduces the required number of annotated spikes for training by 44% compared to training from scratch, achieving up to 25.68% higher accuracy. Additionally, our novel post-processing algorithm is compatible with deep learning frameworks, making E-Sort significantly faster than state-of-the-art spike sorters. On synthesized Neuropixels recordings, E-Sort achieves comparable accuracy with Kilosort4 while sorting 50 seconds of data in only 1.32 seconds. Our method demonstrates robustness across various probe geometries, noise levels, and drift conditions, offering a substantial improvement in both accuracy and runtime efficiency compared to existing spike sorters.

eess.SP↗