arXiv subjects

Zhe Yin

Publications and source records attributed to Zhe Yin.

8 recordsLinked to original sources

arXiv2026

ToolHazard: Scaling Adversarial Environments for Security Evaluation and Alignment of LLM-based Agents

Large language model (LLM) agents integrated with external tools are vulnerable to indirect prompt injections embedded in environmental states. However, existing studies largely rely on manually implemented or reused environments, stochastic LLM-based tool simulation, and predefined injection locations, limiting scalable security research across broader domains. To bridge this gap, we propose **ToolHazard**, a scalable adversarial environment synthesis framework that reduces human engineering and supports expansion with additional seed domains and compute. Through an Environment Simulator, an Attacker Agent, and a User Simulator, ToolHazard synthesizes executable stateful environments, discovers viable injection points and generates environment-specific payloads, and constructs state-grounded long-horizon tasks. Based on ToolHazard, we build **ToolHazard-Bench** for stress-testing agents under complex workflows and diverse environmental attacks. Experiments reveal substantial agent vulnerabilities and show that injection timing and placement affect attack effectiveness. Moreover, ToolHazard-generated alignment data improves security on both ToolHazard-Bench and AgentDojo while preserving benign task utility.

cs.CR↗

arXiv2025

Neuron-Guided Interpretation of Code LLMs: Where, Why, and How?

Code language models excel on code intelligence tasks, yet their internal interpretability is underexplored. Existing neuron interpretability techniques from NLP are suboptimal for source code due to programming languages formal, hierarchical, and executable nature. We empirically investigate code LLMs at the neuron level, localizing language-specific neurons (selectively responsive to one language) and concept layers (feed-forward layers encoding language-agnostic code representations). We analyze Llama-3.1-8B and Qwen2.5-Coder-32B on multilingual inputs in C++, Java, Python, Go, and JavaScript, measuring neuron selectivity and layerwise contributions during generation. We find (1) neurons specialized for individual languages alongside a universal subset supporting general-purpose generation; and (2) lower layers mainly encode language-specific syntax, while middle layers capture semantic abstractions shared across languages, emerging as concept layers. We demonstrate utility on three tasks: neuron-guided fine-tuning for code generation, clone detection via concept-layer embeddings, and concept-layer-guided transfer for code summarization, each yielding consistent gains in multilingual settings.

cs.SE↗

arXiv2025

Multi-Layer Backward Joint Model for Dynamic Prediction of Clinical Events with Multivariate Longitudinal Predictors of Mixed Types

Dynamic prediction of time-to-event outcomes using longitudinal data is highly useful in clinical research and practice. A common strategy is the joint modeling of longitudinal and time-to-event data. The shared random effect model has been widely studied for this purpose. However, it can be computationally challenging when applied to problems with a large number of longitudinal predictor variables, particularly when mixed types of continuous and categorical variables are involved. Addressing these limitations, we introduce a novel multi-layer backward joint model (MBJM). The model structure consists of multiple data layers cohesively integrated through a series of conditional distributions that involve longitudinal and time-to-event data, where the time to the clinical event is the conditioning variable. This model can be estimated with standard statistical software with rapid and robust computation, regardless of the dimension of the longitudinal predictor variables. We provide both theoretical and empirical results to show that the MBJM outperforms the static prediction model that does not fully account for the longitudinal nature of the prediction. In an empirical comparison with the shared random effects joint model, the MBJM demonstrated competitive performance with substantially faster and more robust computation. Both the simulation and real data application from a primary biliary cirrhosis study utilized seven longitudinal biomarkers, five continuous and two categorical, larger than the typically published joint modeling problems.

stat.ME↗

arXiv2025

Anti-adversarial Learning: Desensitizing Prompts for Large Language Models

With the widespread use of LLMs, preserving privacy in user prompts has become crucial, as prompts risk exposing privacy and sensitive data to the cloud LLMs. Traditional techniques like homomorphic encryption, secure multi-party computation, and federated learning face challenges due to heavy computational costs and user participation requirements, limiting their applicability in LLM scenarios. In this paper, we propose PromptObfus, a novel method for desensitizing LLM prompts. The core idea of PromptObfus is "anti-adversarial" learning, which perturbs privacy words in the prompt to obscure sensitive information while retaining the stability of model predictions. Specifically, PromptObfus frames prompt desensitization as a masked language modeling task, replacing privacy-sensitive terms with a [MASK] token. A desensitization model is trained to generate candidate replacements for each masked position. These candidates are subsequently selected based on gradient feedback from a surrogate model, ensuring minimal disruption to the task output. We demonstrate the effectiveness of our approach on three NLP tasks. Results show that PromptObfus effectively prevents privacy inference from remote LLMs while preserving task performance.

cs.CL↗

arXiv2023

Backward Joint Model for the Joint Dynamic Prediction of Time-to-Event and Longitudinal Data: Basic Formulation and New Developments

Dynamic prediction of future clinical outcomes based on longitudinally measured predictors plays a crucial role in disease management and patient counseling, particularly when conventional static models are inadequate. Joint modeling of longitudinal and time-to-event data provides a useful framework for addressing this challenge. In this paper, we present a comprehensive development of the recently proposed backward joint model (BJM; Shen and Li 2021}, which factorizes the likelihood into the distribution of time-to-event data and the conditional distribution of longitudinal data given the event time. This structure facilitates computation and is well-suited for multivariate longitudinal data. We introduce several novel developments to the BJM, including the extrapolation and two-part specifications, as well as the incorporation of competing risks. We also address an important yet underexplored problem in the literature: predicting future longitudinal trajectories conditional on predicted event times. Additionally, we explore the connection between BJM and existing joint modeling approaches. All these extensions preserve the computational advantages of the basic BJM formulation, including one-dimensional numerical integration, convex optimization via the EM algorithm, and a quick procedure for consistent estimation using standard software. We evaluate the method's performance through simulation studies and illustrate its utility in a chronic kidney disease application.

stat.ME↗

arXiv2023

PanoVPR: Towards Unified Perspective-to-Equirectangular Visual Place Recognition via Sliding Windows across the Panoramic View

Visual place recognition has gained significant attention in recent years as a crucial technology in autonomous driving and robotics. Currently, the two main approaches are the perspective view retrieval (P2P) paradigm and the equirectangular image retrieval (E2E) paradigm. However, it is practical and natural to assume that users only have consumer-grade pinhole cameras to obtain query perspective images and retrieve them in panoramic database images from map providers. To address this, we propose \textit{PanoVPR}, a perspective-to-equirectangular (P2E) visual place recognition framework that employs sliding windows to eliminate feature truncation caused by hard cropping. Specifically, PanoVPR slides windows over the entire equirectangular image and computes feature descriptors for each window, which are then compared to determine place similarity. Notably, our unified framework enables direct transfer of the backbone from P2P methods without any modification, supporting not only CNNs but also Transformers. To facilitate training and evaluation, we derive the Pitts250k-P2E dataset from the Pitts250k and establish YQ360, latter is the first P2E visual place recognition dataset collected by a mobile robot platform aiming to simulate real-world task scenarios better. Extensive experiments demonstrate that PanoVPR achieves state-of-the-art performance and obtains 3.8% and 8.0% performance gain on Pitts250k-P2E and YQ360 compared to the previous best method, respectively. Code and datasets will be publicly available at https://github.com/zafirshi/PanoVPR.

cs.CV↗

arXiv2022

PanoFlow: Learning 360° Optical Flow for Surrounding Temporal Understanding

Optical flow estimation is a basic task in self-driving and robotics systems, which enables to temporally interpret traffic scenes. Autonomous vehicles clearly benefit from the ultra-wide Field of View (FoV) offered by 360° panoramic sensors. However, due to the unique imaging process of panoramic cameras, models designed for pinhole images do not directly generalize satisfactorily to 360° panoramic images. In this paper, we put forward a novel network framework--PanoFlow, to learn optical flow for panoramic images. To overcome the distortions introduced by equirectangular projection in panoramic transformation, we design a Flow Distortion Augmentation (FDA) method, which contains radial flow distortion (FDA-R) or equirectangular flow distortion (FDA-E). We further look into the definition and properties of cyclic optical flow for panoramic videos, and hereby propose a Cyclic Flow Estimation (CFE) method by leveraging the cyclicity of spherical images to infer 360° optical flow and converting large displacement to relatively small displacement. PanoFlow is applicable to any existing flow estimation method and benefits from the progress of narrow-FoV flow estimation. In addition, we create and release a synthetic panoramic dataset FlowScape based on CARLA to facilitate training and quantitative analysis. PanoFlow achieves state-of-the-art performance on the public OmniFlowNet and the established FlowScape benchmarks. Our proposed approach reduces the End-Point-Error (EPE) on FlowScape by 27.3%. On OmniFlowNet, PanoFlow achieves a 55.5% error reduction from the best published result. We also qualitatively validate our method via a collection vehicle and a public real-world OmniPhotos dataset, indicating strong potential and robustness for real-world navigation applications. Code and dataset are publicly available at https://github.com/MasterHow/PanoFlow.

cs.CV↗

arXiv2020

A fractional-order SEIHDR model for COVID-19 with inter-city networked coupling effects

In this paper, a mathematical model is proposed to analyze the dynamic behavior of COVID-19. Based on inter-city networked coupling effects, a fractional-order SEIHDR system with the real-data from 23 January to 18 March, 2020 of COVID-19 is discussed. Meanwhile, hospitalized individuals and the mortality rates of three types of individuals (exposed, infected and hospitalized) are firstly taken into account in the proposed model. And infectivity of individuals during incubation is also considered in this paper. By applying least squares method and predictor-correctors scheme, the numerical solutions of the proposed system in the absence of the inter-city network and with the inter-city network are stimulated by using the real-data from 23 January to $18-m$ March, 2020 where $m$ is equal to the number of prediction days. Compared with integer-order system ($α=0$), the fractional-order model without network is validated to have a better fitting of the data on Beijing, Shanghai, Wuhan, Huanggang and other cities. In contrast to the case without network, the results indicate that the inter-city network system may be not a significant case to virus spreading for China because of the lock down and quarantine measures, however, it may have an impact on cities that have not adopted city closure. Meanwhile, the proposed model better fits the data from 24 February to 31, March in Italy, and the peak number of confirmed people is also predicted by this fraction-order model. Furthermore, the existence and uniqueness of a bounded solution under the initial condition are considered in the proposed system. Afterwards, the basic reproduction number $R_0$ is analyzed and it is found to hold a threshold: the disease-free equilibrium point is locally asymptotically stable when $R_0\le 1$, which provides a theoretical basis for whether COVID-19 will become a pandemic in the future.

physics.soc-ph↗