arXiv subjects

Yingjie Zhang

Publications and source records attributed to Yingjie Zhang.

At least 37 records · Page 2Linked to original sources

arXiv2025

A benchmark multimodal oro-dental dataset for large vision-language models

The advancement of artificial intelligence in oral healthcare relies on the availability of large-scale multimodal datasets that capture the complexity of clinical practice. In this paper, we present a comprehensive multimodal dataset, comprising 8775 dental checkups from 4800 patients collected over eight years (2018-2025), with patients ranging from 10 to 90 years of age. The dataset includes 50000 intraoral images, 8056 radiographs, and detailed textual records, including diagnoses, treatment plans, and follow-up notes. The data were collected under standard ethical guidelines and annotated for benchmarking. To demonstrate its utility, we fine-tuned state-of-the-art large vision-language models, Qwen-VL 3B and 7B, and evaluated them on two tasks: classification of six oro-dental anomalies and generation of complete diagnostic reports from multimodal inputs. We compared the fine-tuned models with their base counterparts and GPT-4o. The fine-tuned models achieved substantial gains over these baselines, validating the dataset and underscoring its effectiveness in advancing AI-driven oro-dental healthcare solutions. The dataset is publicly available, providing an essential resource for future research in AI dentistry.

cs.CV↗

arXiv2025

Beyond Surface Alignment: Rebuilding LLMs Safety Mechanism via Probabilistically Ablating Refusal Direction

Jailbreak attacks pose persistent threats to large language models (LLMs). Current safety alignment methods have attempted to address these issues, but they experience two significant limitations: insufficient safety alignment depth and unrobust internal defense mechanisms. These limitations make them vulnerable to adversarial attacks such as prefilling and refusal direction manipulation. We introduce DeepRefusal, a robust safety alignment framework that overcomes these issues. DeepRefusal forces the model to dynamically rebuild its refusal mechanisms from jailbreak states. This is achieved by probabilistically ablating the refusal direction across layers and token depths during fine-tuning. Our method not only defends against prefilling and refusal direction attacks but also demonstrates strong resilience against other unseen jailbreak strategies. Extensive evaluations on four open-source LLM families and six representative attacks show that DeepRefusal reduces attack success rates by approximately 95%, while maintaining model capabilities with minimal performance degradation.

cs.CR↗

arXiv2025

Coexistence of Two Types of Liquid Structures at Platinum-Water Interfaces

Platinum-water interfaces underpin many electrochemical energy conversion processes. However, despite decades of research, the real-space liquid structure of these interfaces remains elusive. Using three-dimensional atomic force microscopy (3D-AFM), we mapped Pt-water interface in real space with angstrom-level resolution. Topographic imaging revealed atomically flat (type I) and stripe-like (type II) surface nanodomains. Force-distance profiles above type I domains exhibited oscillatory decay patterns with periodicity of ~0.33 nm, consistent with water. In contrast, type II domains showed stronger oscillations with larger periodicity of ~0.45 nm and extended decay lengths, indicative of a different liquid structure with stronger correlation and ordering. Wide-angle X-ray scattering (WAXS) measurements of pure water and a series of liquid n-alkanes revealed peaks at ~0.31 nm and ~0.46 nm, in agreement with 3D-AFM observations of type I and type II structures, respectively. Our findings uncover the coexistence of two types of liquid structures at Pt-water interfaces modulated by surface heterogeneity, enabling new understandings and design principles for energy conversion applications.

cond-mat.mtrl-sci↗

arXiv2025

Transformer-based Spatial Grounding: A Comprehensive Survey

Spatial grounding, the process of associating natural language expressions with corresponding image regions, has rapidly advanced due to the introduction of transformer-based models, significantly enhancing multimodal representation and cross-modal alignment. Despite this progress, the field lacks a comprehensive synthesis of current methodologies, dataset usage, evaluation metrics, and industrial applicability. This paper presents a systematic literature review of transformer-based spatial grounding approaches from 2018 to 2025. Our analysis identifies dominant model architectures, prevalent datasets, and widely adopted evaluation metrics, alongside highlighting key methodological trends and best practices. This study provides essential insights and structured guidance for researchers and practitioners, facilitating the development of robust, reliable, and industry-ready transformer-based spatial grounding models.

cs.CV↗

arXiv2025

Correlative angstrom-scale microscopy and spectroscopy of graphite-water interfaces

Water at solid surfaces is key for many processes ranging from biological signal transduction to membrane separation and renewable energy conversion. However, under realistic conditions, which often include environmental and surface charge variations, the interfacial water structure remains elusive. Here we overcome this limit by combining three-dimensional atomic force microscopy and interface-sensitive Raman spectroscopy to characterize the graphite-water interfacial structure in situ. Through correlative analysis of the spatial liquid density maps and vibrational peaks within ~2 nm of the graphite surface, we find the existence of two interfacial configurations at open circuit potential, a transient state where pristine water exhibits strong hydrogen bond (HB) breaking effects, and a steady state with hydrocarbons dominating the interface and weak HB breaking in the surrounding water. At sufficiently negative potentials, both states transition into a stable structure featuring pristine water with a broader distribution of HB configurations. Our three-state model resolves many long-standing controversies on interfacial water structure.

physics.chem-ph↗

arXiv2025

High-Dimensional Dynamic Covariance Models with Random Forests

This paper introduces a novel nonparametric method for estimating high-dimensional dynamic covariance matrices with multiple conditioning covariates, leveraging random forests and supported by robust theoretical guarantees. Unlike traditional static methods, our dynamic nonparametric covariance models effectively capture distributional heterogeneity. Furthermore, unlike kernel-smoothing methods, which are restricted to a single conditioning covariate, our approach accommodates multiple covariates in a fully nonparametric framework. To the best of our knowledge, this is the first method to use random forests for estimating high-dimensional dynamic covariance matrices. In high-dimensional settings, we establish uniform consistency theory, providing nonasymptotic error rates and model selection properties, even when the response dimension grows sub-exponentially with the sample size. These results hold uniformly across a range of conditioning variables. The method's effectiveness is demonstrated through simulations and a stock dataset analysis, highlighting its ability to model complex dynamics in high-dimensional scenarios.

stat.ML↗

arXiv2025

Unraveling Human-AI Teaming: A Review and Outlook

Artificial Intelligence (AI) is advancing at an unprecedented pace, with clear potential to enhance decision-making and productivity. Yet, the collaborative decision-making process between humans and AI remains underdeveloped, often falling short of its transformative possibilities. This paper explores the evolution of AI agents from passive tools to active collaborators in human-AI teams, emphasizing their ability to learn, adapt, and operate autonomously in complex environments. This paradigm shifts challenges traditional team dynamics, requiring new interaction protocols, delegation strategies, and responsibility distribution frameworks. Drawing on Team Situation Awareness (SA) theory, we identify two critical gaps in current human-AI teaming research: the difficulty of aligning AI agents with human values and objectives, and the underutilization of AI's capabilities as genuine team members. Addressing these gaps, we propose a structured research outlook centered on four key aspects of human-AI teaming: formulation, coordination, maintenance, and training. Our framework highlights the importance of shared mental models, trust-building, conflict resolution, and skill adaptation for effective teaming. Furthermore, we discuss the unique challenges posed by varying team compositions, goals, and complexities. This paper provides a foundational agenda for future research and practical design of sustainable, high-performing human-AI teams.

cs.HC↗

arXiv2025

Towards Quantitative Interpretation of 3D Atomic Force Microscopy at Solid-Liquid Interfaces

Three-dimensional atomic force microscopy (3D-AFM) has been a powerful tool to probe the atomic-scale structure of solid-liquid interfaces. As a nanoprobe moves along the 3D volume of interfacial liquid, the probe-sample interaction force is sensed and mapped, providing information on not only the solid morphology, but also the liquid density distribution. To date 3D-AFM force maps of a diverse set of solid-liquid interfaces have been recorded, revealing remarkable force oscillations that are typically attributed to solvation layers or electrical double layers. However, despite the high resolution down to sub-angstrom level, quantitative interpretation of the 3D force maps has been an outstanding challenge. Here we will review the technical details of 3D-AFM and the existing approaches for quantitative data interpretation. Based on evidences in recent literature, we conclude that the perturbation-induced AFM force paradoxically represents the intrinsic, unperturbed liquid density profile. We will further discuss how the oscillatory force profiles can be attributed to the probe-modulation of the liquid configurational entropy, and how the quantitative, atomic-scale liquid density distribution can be derived from the force maps.

physics.chem-ph↗

arXiv2024

Conditional Diffusion Models Based Conditional Independence Testing

Conditional independence (CI) testing is a fundamental task in modern statistics and machine learning. The conditional randomization test (CRT) was recently introduced to test whether two random variables, $X$ and $Y$, are conditionally independent given a potentially high-dimensional set of random variables, $Z$. The CRT operates exceptionally well under the assumption that the conditional distribution $X|Z$ is known. However, since this distribution is typically unknown in practice, accurately approximating it becomes crucial. In this paper, we propose using conditional diffusion models (CDMs) to learn the distribution of $X|Z$. Theoretically and empirically, it is shown that CDMs closely approximate the true conditional distribution. Furthermore, CDMs offer a more accurate approximation of $X|Z$ compared to GANs, potentially leading to a CRT that performs better than those based on GANs. To accommodate complex dependency structures, we utilize a computationally efficient classifier-based conditional mutual information (CMI) estimator as our test statistic. The proposed testing procedure performs effectively without requiring assumptions about specific distribution forms or feature dependencies, and is capable of handling mixed-type conditioning sets that include both continuous and discrete variables. Theoretical analysis shows that our proposed test achieves a valid control of the type I error. A series of experiments on synthetic data demonstrates that our new test effectively controls both type-I and type-II errors, even in high dimensional scenarios.

stat.ML↗

arXiv2024

Bending, breaking, and reconnecting of the electrical double layers at heterogeneous electrodes

In electrochemical systems, the structure of electrical double layers (EDLs) near electrode surfaces is crucial for energy conversion and storage functions. While the electrodes in real-world systems are usually heterogeneous, to date the investigation of EDLs is mainly limited to flat model solid surfaces. To bridge this gap, here we image the EDL structure of an ionic liquid-based electrolyte at a heterogeneous graphite electrode using our recently developed electrochemical 3D atomic force microscopy. These interfaces feature the formation of thin, nanoscale adlayer/cluster domains that closely mimic the early-stage solid-electrolyte interphases in many battery systems. We observe multiple discrete layers in the EDL near the flat electrode, which restructures at the heterogeneous interphase sites. Depending on the local size of the interphase clusters, the EDLs exhibit bending, breaking, and/or reconnecting behaviors, likely due to the combined steric and long-range interaction effects. These results shed light on the fundamental structure and reconfiguration mechanism of EDLs at heterogeneous interfaces.

physics.chem-ph↗

arXiv2024

Endoscopic Fourier-transform infrared spectroscopy through a fiber microprobe

Fourier-transform infrared spectroscopy (FTIR) is a powerful analytical method for not only the chemical identification of solid, liquid, and gas species, but also the quantification of their concentration. However, the chemical quantification capability of FTIR is significantly hindered when the analyte is surrounded by a strong IR absorbing medium, such as liquid solutions. To overcome this limit, here we develop an IR fiber microprobe that can be inserted into liquid medium, and obtain full FTIR spectra at points of interest. To benchmark this endoscopic FTIR method, we insert the microprobe into bulk water covering a ZnSe substrate and measure the IR transmittance of water as a function of the probe-substrate distance. The obtained vibrational modes, overall transmittance vs z profiles, quantitative absorption coefficients, and micro z-section IR transmittance spectra are all consistent with the standard IR absorption properties of water. The results pave the way for endoscopic chemical profiling inside bulk liquid solutions, promising for applications in many biological, chemical, and electrochemical systems.

physics.ins-det↗

arXiv2024

Growth and characterization of the La$_{3}$Ni$_{2}$O$_{7-δ}$ thin films: dominant contribution of the $d_{x^{2}-y^{2}}$ orbital at ambient pressure

By using the pulsed-laser-ablation technique, we have successfully grown the La$_{3}$Ni$_{2}$O$_{7-δ}$ thin films with $c$-axis orientation perpendicular to the film surface. X-ray diffraction shows that the (00l) peaks can be well indexed to the La$_{3}$Ni$_{2}$O$_{7-δ}$ phase. Resistive measurements show that the samples can be tuned from weak insulating to metallic behavior through adjusting the growth conditions. Surprisingly, all curves of $ρ-T$ in the temperature region of 2$\sim$300~K do not show the anomalies corresponding to either the spin density wave or the charge density wave orders as seen in bulk samples. Hall effect measurements show a linear field dependence with the dominant hole charge carriers, but the Hall coefficient $R_{H}=ρ_{xy}/H$ exhibits strong temperature dependence. The magnetoresistance above about 50~K is positive but very weak, indicating a weakened or absence of multiband effect. However, a negative magnetoresistance is observed at low temperatures, which shows the delocalization effect by magnetic field. Detailed analysis on the magnetoresistance suggests that the delocalization effect at low temperatures is due to the Kondo-like effect, rather than the Anderson weak localization. Our transport results suggest that, the electronic conduction is fulfilled by the $d_{x^{2}-y^{2}}$ orbital with holes as the dominant charge carriers, while the interaction through Hund's coupling with the localized $d_{z^{2}}$ orbital plays an important role in the charge dynamics.

cond-mat.supr-con↗

arXiv2024

Global Challenge for Safe and Secure LLMs Track 1

This paper introduces the Global Challenge for Safe and Secure Large Language Models (LLMs), a pioneering initiative organized by AI Singapore (AISG) and the CyberSG R&D Programme Office (CRPO) to foster the development of advanced defense mechanisms against automated jailbreaking attacks. With the increasing integration of LLMs in critical sectors such as healthcare, finance, and public administration, ensuring these models are resilient to adversarial attacks is vital for preventing misuse and upholding ethical standards. This competition focused on two distinct tracks designed to evaluate and enhance the robustness of LLM security frameworks. Track 1 tasked participants with developing automated methods to probe LLM vulnerabilities by eliciting undesirable responses, effectively testing the limits of existing safety protocols within LLMs. Participants were challenged to devise techniques that could bypass content safeguards across a diverse array of scenarios, from offensive language to misinformation and illegal activities. Through this process, Track 1 aimed to deepen the understanding of LLM vulnerabilities and provide insights for creating more resilient models.

cs.CR↗

arXiv2024

Enhancing Missing Data Imputation through Combined Bipartite Graph and Complete Directed Graph

In this paper, we aim to address a significant challenge in the field of missing data imputation: identifying and leveraging the interdependencies among features to enhance missing data imputation for tabular data. We introduce a novel framework named the Bipartite and Complete Directed Graph Neural Network (BCGNN). Within BCGNN, observations and features are differentiated as two distinct node types, and the values of observed features are converted into attributed edges linking them. The bipartite segment of our framework inductively learns embedding representations for nodes, efficiently utilizing the comprehensive information encapsulated in the attributed edges. In parallel, the complete directed graph segment adeptly outlines and communicates the complex interdependencies among features. When compared to contemporary leading imputation methodologies, BCGNN consistently outperforms them, achieving a noteworthy average reduction of 15% in mean absolute error for feature imputation tasks under different missing mechanisms. Our extensive experimental investigation confirms that an in-depth grasp of the interdependence structure substantially enhances the model's feature embedding ability. We also highlight the model's superior performance in label prediction tasks involving missing data, and its formidable ability to generalize to unseen data points.

cs.LG↗

arXiv2024

Integrating Experiment with Theory to Determine the Structure of Electrode-Electrolyte Interfaces

Electrode-electrolyte interfaces are crucial for electrochemical energy conversion and storage. At these interfaces, the liquid electrolytes form electrical double layers (EDLs). However, despite more than a century of active research, the fundamental structure of EDLs remains elusive to date. Experimental characterization and theoretical calculations have both provided insights, yet each method by itself only offers incomplete or inexact information of the multifaceted EDL structure. Here we provide a survey of the mainstream approaches for EDL quantification, with a particular focus on the emerging 3D atomic force microscopy (3D-AFM) imaging which provides real-space atomic-scale EDL structures. To overcome the existing limits of EDL characterization methods, we propose a new approach to integrate 3D-AFM with classical molecular dynamics (MD) simulation, to enable realistic, precise, and high-throughput determination and prediction of EDL structures. As examples of real-world application, we will discuss the feasibility of using this joint experiment-theory method to unravel the EDL structure at various carbon-based electrodes for supercapacitors, batteries, and electrocatalysis. Looking forward, we believe 3D-AFM, future versions of scanning probe microscopy, and their integration with theory offer promising platforms to profile liquid structures in many electrochemical systems.

physics.chem-ph↗

arXiv2024

Making Them Ask and Answer: Jailbreaking Large Language Models in Few Queries via Disguise and Reconstruction

In recent years, large language models (LLMs) have demonstrated notable success across various tasks, but the trustworthiness of LLMs is still an open problem. One specific threat is the potential to generate toxic or harmful responses. Attackers can craft adversarial prompts that induce harmful responses from LLMs. In this work, we pioneer a theoretical foundation in LLMs security by identifying bias vulnerabilities within the safety fine-tuning and design a black-box jailbreak method named DRA (Disguise and Reconstruction Attack), which conceals harmful instructions through disguise and prompts the model to reconstruct the original harmful instruction within its completion. We evaluate DRA across various open-source and closed-source models, showcasing state-of-the-art jailbreak success rates and attack efficiency. Notably, DRA boasts a 91.1% attack success rate on OpenAI GPT-4 chatbot.

cs.CR↗

arXiv2024

Nodewise Loreg: Nodewise $L_0$-penalized Regression for High-dimensional Sparse Precision Matrix Estimation

We propose Nodewise Loreg, a nodewise $L_0$-penalized regression method for estimating high-dimensional sparse precision matrices. We establish its asymptotic properties, including convergence rates, support recovery, and asymptotic normality under high-dimensional sub-Gaussian settings. Notably, the Nodewise Loreg estimator is asymptotically unbiased and normally distributed, eliminating the need for debiasing required by Nodewise Lasso. We also develop a desparsified version of Nodewise Loreg, similar to the desparsified Nodewise Lasso estimator. The asymptotic variances of the undesparsified Nodewise Loreg estimator are upper bounded by those of both desparsified Nodewise Loreg and Lasso estimators for Gaussian data, potentially offering more powerful statistical inference. Extensive simulations show that the undesparsified Nodewise Loreg estimator generally outperforms the two desparsified estimators in asymptotic normal behavior. Moreover, Nodewise Loreg surpasses Nodewise Lasso, CLIME, and GLasso in most simulations in terms of matrix norm losses, support recovery, and timing performance. Application to a breast cancer gene expression dataset further demonstrates Nodewise Loreg's superiority over the three $L_1$-norm based methods.

math.ST↗

arXiv2024

Confidence-Aware RGB-D Face Recognition via Virtual Depth Synthesis

2D face recognition encounters challenges in unconstrained environments due to varying illumination, occlusion, and pose. Recent studies focus on RGB-D face recognition to improve robustness by incorporating depth information. However, collecting sufficient paired RGB-D training data is expensive and time-consuming, hindering wide deployment. In this work, we first construct a diverse depth dataset generated by 3D Morphable Models for depth model pre-training. Then, we propose a domain-independent pre-training framework that utilizes readily available pre-trained RGB and depth models to separately perform face recognition without needing additional paired data for retraining. To seamlessly integrate the two distinct networks and harness the complementary benefits of RGB and depth information for improved accuracy, we propose an innovative Adaptive Confidence Weighting (ACW). This mechanism is designed to learn confidence estimates for each modality to achieve modality fusion at the score level. Our method is simple and lightweight, only requiring ACW training beyond the backbone models. Experiments on multiple public RGB-D face recognition benchmarks demonstrate state-of-the-art performance surpassing previous methods based on depth estimation and feature fusion, validating the efficacy of our approach.

cs.CV↗