Haojun Wang | Searcharxiv

arXiv2026

Workflow-GYM: Towards Long-Horizon Evaluation of Computer-use Agentic tasks in Real-World Professional Fields

Recent years have witnessed the rapid evolution of AI agents toward handling increasingly complex, real-world tasks. However, existing benchmarks rarely evaluate whether agents can operate graphical user interfaces to complete long-horizon, high-value professional workflows across diverse domains. Current GUI benchmarks still predominantly focus on general-purpose software, relatively simple applications, and short-horizon tasks, leaving it largely unknown whether modern agents can follow user instructions to autonomously operate domain-specific professional software and accomplish economically valuable work in an end-to-end manner. To bridge this gap, we introduce Workflow-GYM, a benchmark for long-horizon GUI tasks centered on professional domains and specialized software environments. Through extensive experiments on state-of-the-art models, we find that even the strongest models achieve only slightly above 30% success rates, highlighting that professional long-horizon GUI workflows remain highly challenging for current GUI agents. Further analysis reveals that current agents struggle to maintain long-horizon workflow consistency, frequently exhibiting workflow stage omission, error propagation, objective drift, and insufficient understanding of professional software environments. Our findings provide important insights into the limitations of current agent systems and suggest key directions for the next generation of GUI-agent research.

cs.AI↗

arXiv2024

Privacy and Security Trade-off in Interconnected Systems with Known or Unknown Privacy Noise Covariance

This paper is concerned with the security problem for interconnected systems, where each subsystem is required to detect local attacks using locally available information and the information received from its neighboring subsystems. Moreover, we consider that there exists an additional eavesdropper being able to infer the private information by eavesdropping transmitted data between subsystems. Then, a privacy-preserving method is employed by adding privacy noise to transmitted data, and the privacy level is measured by mutual information. Nevertheless, adding privacy noise to transmitted data may affect the detection performance metrics such as detection probability and false alarm probability. Thus, we theoretically analyze the trade-off between the privacy and the detection performance. An optimization problem with maximizing both the degree of privacy preservation and the detection probability is established to obtain the covariance of the privacy noise. In addition, the attack detector of each subsystem may not obtain all information about the privacy noise. We further theoretically analyze the trade-off between the privacy and the false alarm probability when the attack detector has no knowledge of the privacy noise covariance. An optimization problem with maximizing the degree of privacy preservation with guaranteeing a bound of false alarm distortion level is established to obtain {\color{black}{the covariance of the privacy noise}}. Moreover, to analyze the effect of the privacy noise on the detection probability, we consider that each subsystem can estimate the unknown privacy noise covariance by the secondary data. Based on the estimated covariance, we construct another attack detector and analyze how the privacy noise affects its detection performance. Finally, a numerical example is provided to verify the effectiveness of theoretical results.

eess.SY↗