arXiv subjects

Xiaoyu Luo

Publications and source records attributed to Xiaoyu Luo.

At least 19 recordsLinked to original sources

arXiv2026

Characterizing Memorization in Diffusion Language Models: Generalized Extraction and Sampling Effects

Autoregressive language models (ARMs) have been shown to memorize and occasionally reproduce training data verbatim, raising concerns about privacy and copyright liability. Diffusion language models (DLMs) have recently emerged as a competitive alternative, yet their memorization behavior remains largely unexplored due to fundamental differences in generation dynamics. To address this gap, we present a systematic theoretical and empirical characterization of memorization in DLMs. We propose a generalized probabilistic extraction framework that unifies prefix-conditioned decoding and diffusion-based generation under arbitrary masking patterns and stochastic sampling trajectories. Theorem 4.3 establishes a monotonic relationship between sampling resolution and memorization: increasing resolution strictly increases the probability of exact training data extraction, implying that autoregressive decoding corresponds to a limiting case of diffusion-based generation by setting the sampling resolution maximal. Extensive experiments across model scales and sampling strategies validate our theoretical predictions. Under aligned prefix-conditioned evaluations, we further demonstrate that DLMs exhibit substantially lower memorization-based leakage of personally identifiable information (PII) compared to ARMs.

cs.CL↗

arXiv2026

InsightX Agent: An LMM-based Agentic Framework with Integrated Tools for Reliable X-ray NDT Analysis

Non-destructive testing (NDT), particularly X-ray inspection, is vital for industrial quality assurance, yet existing deep-learning-based approaches often lack interactivity, interpretability, and the capacity for critical self-assessment, limiting their reliability and operator trust. To address these shortcomings, this paper proposes InsightX Agent, a novel LMM-based agentic framework designed to deliver reliable, interpretable, and interactive X-ray NDT analysis. Unlike typical sequential pipelines, InsightX Agent positions a Large Multimodal Model (LMM) as a central orchestrator, coordinating between the Sparse Deformable Multi-Scale Detector (SDMSD) and the Evidence-Grounded Reflection (EGR) tool. The SDMSD generates dense defect region proposals from multi-scale feature maps and sparsifies them through Non-Maximum Suppression (NMS), optimizing detection of small, dense targets in X-ray images while maintaining computational efficiency. The EGR tool guides the LMM agent through a chain-of-thought-inspired review process, incorporating context assessment, individual defect analysis, false positive elimination, confidence recalibration and quality assurance to validate and refine the SDMSD's initial proposals. By strategically employing and intelligently using tools, InsightX Agent moves beyond passive data processing to active reasoning, enhancing diagnostic reliability and providing interpretations that integrate diverse information sources. Experimental evaluations on the GDXray+ dataset demonstrate that InsightX Agent not only achieves a high object detection F1-score of 96.54\% but also offers significantly improved interpretability and trustworthiness in its analyses, highlighting the transformative potential of LMM-based agentic frameworks for industrial inspection tasks.

cs.AI↗

arXiv2026

APEX: Probing Neural Networks via Activation Perturbation

Prior work on probing neural networks primarily relies on input-space analysis or parameter perturbation, both of which face fundamental limitations in accessing structural information encoded in intermediate representations. We introduce Activation Perturbation for EXploration (APEX), an inference-time probing paradigm that perturbs hidden activations while keeping both inputs and model parameters fixed. We theoretically show that activation perturbation induces a principled transition from sample-dependent to model-dependent behavior by suppressing input-specific signals and amplifying representation-level structure, and further establish that input perturbation corresponds to a constrained special case of this framework. Through representative case studies, we demonstrate the practical advantages of APEX. In the small-noise regime, APEX provides a lightweight and efficient measure of sample regularity that aligns with established metrics, while also distinguishing structured from randomly labeled models and revealing semantically coherent prediction transitions. In the large-noise regime, APEX exposes training-induced model-level biases, including a pronounced concentration of predictions on the target class in backdoored models. Overall, our results show that APEX offers an effective perspective for exploring, and understanding neural networks beyond what is accessible from input space alone.

cs.LG↗

arXiv2026

Do LLMs Really Memorize Personally Identifiable Information? Revisiting PII Leakage with a Cue-Controlled Memorization Framework

Large Language Models (LLMs) have been reported to "leak" Personally Identifiable Information (PII), with successful PII reconstruction often interpreted as evidence of memorization. We propose a principled revision of memorization evaluation for LLMs, arguing that PII leakage should be evaluated under low lexical cue conditions, where target PII cannot be reconstructed through prompt-induced generalization or pattern completion. We formalize Cue-Resistant Memorization (CRM) as a cue-controlled evaluation framework and a necessary condition for valid memorization evaluation, explicitly conditioning on prompt-target overlap cues. Using CRM, we conduct a large-scale multilingual re-evaluation of PII leakage across 32 languages and multiple memorization paradigms. Revisiting reconstruction-based settings, including verbatim prefix-suffix completion and associative reconstruction, we find that their apparent effectiveness is driven primarily by direct surface-form cues rather than by true memorization. When such cues are controlled for, reconstruction success diminishes substantially. We further examine cue-free generation and membership inference, both of which exhibit extremely low true positive rates. Overall, our results suggest that previously reported PII leakage is better explained by cue-driven behavior than by genuine memorization, highlighting the importance of cue-controlled evaluation for reliably quantifying privacy-relevant memorization in LLMs.

cs.CL↗

arXiv2026

Shared Path: Unraveling Memorization in Multilingual LLMs through Language Similarities

We present the first comprehensive study of Memorization in Multilingual Large Language Models (MLLMs), analyzing 95 languages using models across diverse model scales, architectures, and memorization definitions. As MLLMs are increasingly deployed, understanding their memorization behavior has become critical. Yet prior work has focused primarily on monolingual models, leaving multilingual memorization underexplored, despite the inherently long-tailed nature of training corpora. We find that the prevailing assumption, that memorization is highly correlated with training data availability, fails to fully explain memorization patterns in MLLMs. We hypothesize that the conventional focus on monolingual settings, effectively treating languages in isolation, may obscure the true patterns of memorization. To address this, we propose a novel graph-based correlation metric that incorporates language similarity to analyze cross-lingual memorization. Our analysis reveals that among similar languages, those with fewer training tokens tend to exhibit higher memorization, a trend that only emerges when cross-lingual relationships are explicitly modeled. These findings underscore the importance of a \textit{language-aware} perspective in evaluating and mitigating memorization vulnerabilities in MLLMs. This also constitutes empirical evidence that language similarity both explains Memorization in MLLMs and underpins Cross-lingual Transferability, with broad implications for multilingual NLP.

cs.CL↗

arXiv2025

Investigations of MWISP Clumps: 13CO Clump Source Catalog and Physical Properties

We present the first comprehensive catalogs of $^{13}$CO clumps from the Milky Way Imaging Scroll Painting (MWISP) project. By developing an equivalent global detection scheme integrated with the FacetClumps algorithm, we successfully extract 71,661 molecular clumps across a high-resolution $^{13}$CO data cube spanning 2310 deg$^2$ from the MWISP Phase I survey. To determine accurate distances, we design an automatic hierarchical distance decision method using signal regions as fundamental objects, effectively resolving the kinematic distance ambiguity problem and obtaining reliable measurements for 97.94% of the sample. Statistical analysis reveals that 65.3% of clumps are gravitationally bound, accounting for approximately 96.3% of the statistical total mass. Scaling relation analysis across multiple surveys reveals universal power-law behaviors in clump populations. Maser-associated clumps exhibit modified parameter distributions and scaling relations, revealing how active star formation alters clump dynamics and structure. These extensive catalogs establish a foundation for investigating molecular clump properties, star formation processes, and Galactic evolution.

astro-ph.GA↗

arXiv2025

Sparse Convex Quantile Regression: A Generalized Benders Decomposition Approach

We develop a scalable algorithmic framework for sparse convex quantile regression (SCQR), addressing key computational challenges in the literature. Enhancing the classical CQR model, we introduce L2-norm regularization and an epsilon-insensitive zone to improve generalization and mitigate overfitting - both theoretically justified and empirically validated. Based on this extension, we improve the SCQR model and propose the first Generalized Benders Decomposition (GBD) algorithm tailored to this context, further strengthened by a novel local search-based Benders matheuristic. Extensive simulations and a real-world application to Sustainable Development Goals benchmarking demonstrate the accuracy, scalability, and practical value of our approach.

math.OC↗

arXiv2025

DeMem: Privacy-Enhanced Robust Adversarial Learning via De-Memorization

Adversarial robustness, the ability of a model to withstand manipulated inputs that cause errors, is essential for ensuring the trustworthiness of machine learning models in real-world applications. However, previous studies have shown that enhancing adversarial robustness through adversarial training increases vulnerability to privacy attacks. While differential privacy can mitigate these attacks, it often compromises robustness against both natural and adversarial samples. Our analysis reveals that differential privacy disproportionately impacts low-risk samples, causing an unintended performance drop. To address this, we propose DeMem, which selectively targets high-risk samples, achieving a better balance between privacy protection and model robustness. DeMem is versatile and can be seamlessly integrated into various adversarial training techniques. Extensive evaluations across multiple training methods and datasets demonstrate that DeMem significantly reduces privacy leakage while maintaining robustness against both natural and adversarial samples. These results confirm DeMem's effectiveness and broad applicability in enhancing privacy without compromising robustness.

cs.LG↗

arXiv2025

Trustworthy Machine Learning via Memorization and the Granular Long-Tail: A Survey on Interactions, Tradeoffs, and Beyond

The role of memorization in machine learning (ML) has garnered significant attention, particularly as modern models are empirically observed to memorize fragments of training data. Previous theoretical analyses, such as Feldman's seminal work, attribute memorization to the prevalence of long-tail distributions in training data, proving it unavoidable for samples that lie in the tail of the distribution. However, the intersection of memorization and trustworthy ML research reveals critical gaps. While prior research in memorization in trustworthy ML has solely focused on class imbalance, recent work starts to differentiate class-level rarity from atypical samples, which are valid and rare intra-class instances. However, a critical research gap remains: current frameworks conflate atypical samples with noisy and erroneous data, neglecting their divergent impacts on fairness, robustness, and privacy. In this work, we conduct a thorough survey of existing research and their findings on trustworthy ML and the role of memorization. More and beyond, we identify and highlight uncharted gaps and propose new revenues in this research direction. Since existing theoretical and empirical analyses lack the nuances to disentangle memorization's duality as both a necessity and a liability, we formalize three-level long-tail granularity - class imbalance, atypicality, and noise - to reveal how current frameworks misapply these levels, perpetuating flawed solutions. By systematizing this granularity, we draw a roadmap for future research. Trustworthy ML must reconcile the nuanced trade-offs between memorizing atypicality for fairness assurance and suppressing noise for robustness and privacy guarantee. Redefining memorization via this granularity reshapes the theoretical foundation for trustworthy ML, and further affords an empirical prerequisite for models that align performance with societal trust.

cs.LG↗

arXiv2024

The Empirical Impact of Forgetting and Transfer in Continual Visual Odometry

As robotics continues to advance, the need for adaptive and continuously-learning embodied agents increases, particularly in the realm of assistance robotics. Quick adaptability and long-term information retention are essential to operate in dynamic environments typical of humans' everyday lives. A lifelong learning paradigm is thus required, but it is scarcely addressed by current robotics literature. This study empirically investigates the impact of catastrophic forgetting and the effectiveness of knowledge transfer in neural networks trained continuously in an embodied setting. We focus on the task of visual odometry, which holds primary importance for embodied agents in enabling their self-localization. We experiment on the simple continual scenario of discrete transitions between indoor locations, akin to a robot navigating different apartments. In this regime, we observe initial satisfactory performance with high transferability between environments, followed by a specialization phase where the model prioritizes current environment-specific knowledge at the expense of generalization. Conventional regularization strategies and increased model capacity prove ineffective in mitigating this phenomenon. Rehearsal is instead mildly beneficial but with the addition of a substantial memory cost. Incorporating action information, as commonly done in embodied settings, facilitates quicker convergence but exacerbates specialization, making the model overly reliant on its motion expectations and less adept at correctly interpreting visual cues. These findings emphasize the open challenges of balancing adaptation and memory retention in lifelong robotics and contribute valuable insights into the application of a lifelong paradigm on embodied agents.

cs.CV↗

arXiv2023

ADPBA: Efficiently generating Lagrangian cuts for two-stage stochastic integer programs

The use of Lagrangian cuts proves effective in enhancing the lower bound of the master problem within the execution of benders-type algorithms, particularly in the context of two-stage stochastic programs. However, even the process of generating a single Lagrangian cut is notably time-intensive. In light of this challenge, we present a novel framework that integrates Lagrangian cut generation with an adaptive partition-based approach, thereby mitigating this time-related drawback to a considerable extent. Furthermore, we also discuss the dominance relationship between the generated partition-based Lagrangian cut and the Lagrangian cut for the original problem. To provide empirical evidence of our approach's efficacy, we undertake an extensive computational study encompassing instances involving even up to a thousand scenarios. The results of this study conclusively demonstrate the superiority and efficiency of the proposed methodology.

math.OC↗

arXiv2023

Heuristic Learning for Co-Design Scheme of Optimal Sequential Attack

This paper considers a novel co-design problem of the optimal \textit{sequential} attack, whose attack strategy changes with the time series, and in which the \textit{sequential} attack selection strategy and \textit{sequential} attack signal are simultaneously designed. Different from the existing attack design works that separately focus on attack subsets or attack signals, the joint design of the attack strategy poses a huge challenge due to the deep coupling relation between the \textit{sequential} attack selection strategy and \textit{sequential} attack signal. In this manuscript, we decompose the sequential co-design problem into two equivalent sub-problems. Specifically, we first derive an analytical closed-form expression between the optimal attack signal and the sequential attack selection strategy. Furthermore, we prove the finite-time inverse convergence of the critical parameters in the injected optimal attack signal by discrete-time Lyapunov analysis, which enables the efficient off-line design of the attack signal and saves computing resources. Finally, we exploit its relationship to design a heuristic two-stage learning-based joint attack algorithm (HTL-JA), which can accelerate realization of the attack target compared to the one-stage proximal-policy-optimization-based (PPO) algorithm. Extensive simulations are conducted to show the effectiveness of the injected optimal sequential attack.

math.OC↗

arXiv2023

Model-free False Data Injection Attack in Networked Control Systems: A Feedback Optimization Approach

Security issues have gathered growing interest within the control systems community, as physical components and communication networks are increasingly vulnerable to cyber attacks. In this context, recent literature has studied increasingly sophisticated \emph{false data injection} attacks, with the aim to design mitigative measures that improve the systems' security. Notably, data-driven attack strategies -- whereby the system dynamics is oblivious to the adversary -- have received increasing attention. However, many of the existing works on the topic rely on the implicit assumption of linear system dynamics, significantly limiting their scope. Contrary to that, in this work we design and analyze \emph{truly} model-free false data injection attack that applies to general linear and nonlinear systems. More specifically, we aim at designing an injected signal that steers the output of the system toward a (maliciously chosen) trajectory. We do so by designing a zeroth-order feedback optimization policy and jointly use probing signals for real-time measurements. We then characterize the quality of the proposed model-free attack through its optimality gap, which is affected by the dimensions of the attack signal, the number of iterations performed, and the convergence rate of the system. Finally, we extend the proposed attack scheme to the systems with internal noise. Extensive simulations show the effectiveness of the proposed attack scheme.

math.OC↗

arXiv2023

FacetClumps: A Facet-based Molecular Clump Detection Algorithm

A comprehensive understanding of molecular clumps is essential for investigating star formation. We present an algorithm for molecular clump detection, called FacetClumps. This algorithm uses a morphological approach to extract signal regions from the original data. The Gaussian Facet model is employed to fit the signal regions, which enhances the resistance to noise and the stability of the algorithm in diverse overlapping areas. The introduction of the extremum determination theorem of multivariate functions offers theoretical guidance for automatically locating clump centers. To guarantee that each clump is continuous, the signal regions are segmented into local regions based on gradient, and then the local regions are clustered into the clump centers based on connectivity and minimum distance to identify the regional information of each clump. Experiments conducted with both simulated and synthetic data demonstrate that FacetClumps exhibits great recall and precision rates, small location error and flux loss, a high consistency between the region of detected clump and that of simulated clump, and is generally stable in various environments. Notably, the recall rate of FacetClumps in the synthetic data, which comprises $^{13}CO$ ($J = 1-0$) emission line of the MWISP within $11.7^{\circ} \leq l \leq 13.4^{\circ}$, $0.22^{\circ} \leq b \leq 1.05^{\circ}$ and 5 km s$^{-1}$ $\leq v \leq$ 35 km s$^{-1}$ and simulated clumps, reaches 90.2%. Additionally, FacetClumps demonstrates satisfactory performance when applied to observational data.

astro-ph.IM↗

arXiv2022

Nonlinear Indentation of Second-order Hyperelastic Materials

The classical problem of indentation on an elastic substrate has found new applications in the field of the Atomic Force Microscopy. However, linearly elastic indentation models are not sufficiently accurate to predict the force-displacement relationship at large indentation depths. For hyperelastic materials, such as soft polymers and biomaterials, a nonlinear indentation model is needed. In this paper, we use second-order elasticity theory to capture larger amplitude deformations and material nonlinearity. We provide a general solution for the contact problem for deformations that are second-order in indentation amplitude with arbitrary indenter profiles. Moreover, we derive analytical solutions by using either parabolic or quartic surfaces to mimic a spherical indenter. The analytical prediction for a quartic surface agrees well with finite element simulations using a spherical indenter for indentation depths on the order of the indenter radius. In particular, the relative error between the two approaches is less than 1% for an indentation depth equal to the indenter radius, an order of magnitude less than that observed with models which are either first-order in indentation amplitude or those which are second-order in indentation amplitude but with a parabolic indenter profile.

cond-mat.soft↗

arXiv2022

Submodularity-based False Data Injection Attack Scheme in Multi-agent Dynamical Systems

Consensus in multi-agent dynamical systems is prone to be sabotaged by the adversary, which has attracted much attention due to its key role in broad applications. In this paper, we study a new false data injection (FDI) attack design problem, where the adversary with limited capability aims to select a subset of agents and manipulate their local multi-dimensional states to maximize the consensus convergence error. We first formulate the FDI attack design problem as a combinatorial optimization problem and prove it is NP-hard. Then, based on the submodularity optimization theory, we show the convergence error is a submodular function of the set of the compromised agents, which satisfies the property of diminishing marginal returns. In other words, the benefit of adding an extra agent to the compromised set decreases as that set becomes larger. With this property, we exploit the greedy scheme to find the optimal compromised agent set that can produce the maximum convergence error when adding one extra agent to that set each time. Thus, the FDI attack set selection algorithms are developed to obtain the near-optimal subset of the compromised agents. Furthermore, we derive the analytical suboptimality bounds and the worst-case running time under the proposed algorithms. Extensive simulation results are conducted to show the effectiveness of the proposed algorithm.

math.DS↗

arXiv2021

Molecular Clump Extraction Algorithm Based on Local Density Clustering

The detection and parametrization of molecular clumps is the first step in studying them. We propose a method based on Local Density Clustering algorithm while physical parameters of those clumps are measured using the Multiple Gaussian Model algorithm. One advantage of applying the Local Density Clustering to the clump detection and segmentation, is the high accuracy under different signal-to-noise levels. The Multiple Gaussian Model is able to deal with overlapping clumps whose parameters can be derived reliably. Using simulation and synthetic data, we have verified that the proposed algorithm could characterize the morphology and flux of molecular clumps accurately. The total flux recovery rate in $^{13}\rm CO$ (J=1-0) line of M16 is measured as 90.2\%. The detection rate and the completeness limit are 81.7\% and 20 K km s$ ^{-1} $ in $^{13}\rm CO$ (J=1-0) line of M16, respectively.

astro-ph.IM↗

arXiv2021

Multiple Steady and Oscillatory Solutions in a Collapsible Channel Flow

We study flow driven through a finite-length planar rigid channel by a fixed upstream flux, where a segment of one wall is replaced by a pre-stressed elastic beam subject to uniform external pressure. The steady and unsteady systems are solved using a finite element method. Previous studies have shown that the system can exhibit three steady states for some parameters (termed the upper, intermediate and lower steady branches, respectively). Of these, the intermediate branch is always unstable while the upper and lower steady branches can (independently) become unstable to self-excited oscillations. We show that for some parameter combinations the system is unstable to both upper and lower branch oscillations simultaneously. However, we show that these two instabilities eventually merge together for large enough Reynolds numbers, exhibiting a nonlinear limit cycle which retains characteristics of both the upper and lower branches of oscillations. Furthermore, we show that increasing the beam pre-tension suppresses the region of multiple steady states but preserves the onset of oscillations. Conversely, increasing the beam thickness (a proxy for increasing bending stiffness) suppresses both multiple steady states and the onset of oscillations.

physics.flu-dyn↗