ASLEval: Measuring Privacy Exposure Displacement in LLM Agent Sessions
本文提出ASLEval框架，通过预注册隐藏目标集、测量所有可见出口并保留内部痕迹来解决多步骤会话中隐私暴露错位的问题。
0 citationsRead paper
Selected work
Representative Papers
Stochasticity Is Not the Hard Part: Reduction and Complexity in Instructional Sequencing over Prerequisite DAGs
This study addresses the problem of finding the optimal teaching sequence that minimizes learning cost in scenarios with prerequisite dependencies. The problem is modeled as a stochastic shortest path problem, and we propose an exact reduction based on lattice theory that transforms it into a deterministic shortest path problem, revealing that the computational difficulty stems not from stochasticity but from the combinatorial complexity inherent in the dependency structure. We theoretically prove the problem to be NP-hard; however, by leveraging dynamic programming, A* search, and feedback arc set reductions, we identify a “doubly simple” regime in real-world course data where A* efficiently solves instances with state-space size linear in the number of concepts. A computable diagnostic metric, \( m\Delta \), further enables practical assessment of instance hardness.
0 citationsRead paper
Glass Surface Segmentation with an RGB-D Camera via Weighted Feature Fusion for Service Robots
Service robots face significant challenges in segmenting glass surfaces using RGB-D cameras in real-world scenarios due to glass transparency, strong reflections, and occlusions. To address these issues, this paper proposes a Weighted Feature Fusion (WFF) module that enables dynamic, adaptive fusion of RGB and depth features; the module is plug-and-play and compatible with multiple mainstream segmentation backbones. Furthermore, we introduce MJU-Glass—the first real-world glass segmentation dataset collected *in situ* by service robots—filling a critical gap in publicly available glass segmentation data. Integrating WFF into architectures such as PSPNet yields substantial improvements in segmentation robustness without compromising computational efficiency: boundary IoU increases by 7.49%, and mean IoU also improves significantly, thereby effectively reducing robot collision risk during navigation and interaction.
0 citationsRead paper
LSU-Net: Lightweight Automatic Organs Segmentation Network For Medical Images
To address the challenge of balancing segmentation accuracy and computational efficiency for organ segmentation in resource-constrained clinical settings (e.g., standard hospital PCs), this paper proposes LSU-Net, a lightweight deep learning architecture. Methodologically, it introduces two key innovations: (i) a novel collaborative block design—Light Conv Block and Tokenized Shift Block—that integrates depthwise separable convolutions with spatial shift mechanisms to drastically reduce parameter count; and (ii) a dynamic weighted multi-task loss function to enhance multi-scale feature representation. Evaluated on the UWMGI and MSD Colon datasets, LSU-Net achieves competitive or superior Dice scores (1.3–2.1% improvement) with only ~1.2M parameters—over 50% fewer than mainstream state-of-the-art models—while accelerating inference by 2.4×. The model thus achieves an optimal trade-off among lightness, accuracy, and deployability, offering a practical paradigm for AI deployment in resource-limited healthcare environments.
0 citationsRead paper
Recent publications
Latest Papers
ASLEval: Measuring Privacy Exposure Displacement in LLM Agent Sessions
本文提出ASLEval框架，通过预注册隐藏目标集、测量所有可见出口并保留内部痕迹来解决多步骤会话中隐私暴露错位的问题。
0 citationsRead paper
Stochasticity Is Not the Hard Part: Reduction and Complexity in Instructional Sequencing over Prerequisite DAGs
This study addresses the problem of finding the optimal teaching sequence that minimizes learning cost in scenarios with prerequisite dependencies. The problem is modeled as a stochastic shortest path problem, and we propose an exact reduction based on lattice theory that transforms it into a deterministic shortest path problem, revealing that the computational difficulty stems not from stochasticity but from the combinatorial complexity inherent in the dependency structure. We theoretically prove the problem to be NP-hard; however, by leveraging dynamic programming, A* search, and feedback arc set reductions, we identify a “doubly simple” regime in real-world course data where A* efficiently solves instances with state-space size linear in the number of concepts. A computable diagnostic metric, \( m\Delta \), further enables practical assessment of instance hardness.
0 citationsRead paper
Glass Surface Segmentation with an RGB-D Camera via Weighted Feature Fusion for Service Robots
Service robots face significant challenges in segmenting glass surfaces using RGB-D cameras in real-world scenarios due to glass transparency, strong reflections, and occlusions. To address these issues, this paper proposes a Weighted Feature Fusion (WFF) module that enables dynamic, adaptive fusion of RGB and depth features; the module is plug-and-play and compatible with multiple mainstream segmentation backbones. Furthermore, we introduce MJU-Glass—the first real-world glass segmentation dataset collected *in situ* by service robots—filling a critical gap in publicly available glass segmentation data. Integrating WFF into architectures such as PSPNet yields substantial improvements in segmentation robustness without compromising computational efficiency: boundary IoU increases by 7.49%, and mean IoU also improves significantly, thereby effectively reducing robot collision risk during navigation and interaction.
0 citationsRead paper
LSU-Net: Lightweight Automatic Organs Segmentation Network For Medical Images
To address the challenge of balancing segmentation accuracy and computational efficiency for organ segmentation in resource-constrained clinical settings (e.g., standard hospital PCs), this paper proposes LSU-Net, a lightweight deep learning architecture. Methodologically, it introduces two key innovations: (i) a novel collaborative block design—Light Conv Block and Tokenized Shift Block—that integrates depthwise separable convolutions with spatial shift mechanisms to drastically reduce parameter count; and (ii) a dynamic weighted multi-task loss function to enhance multi-scale feature representation. Evaluated on the UWMGI and MSD Colon datasets, LSU-Net achieves competitive or superior Dice scores (1.3–2.1% improvement) with only ~1.2M parameters—over 50% fewer than mainstream state-of-the-art models—while accelerating inference by 2.4×. The model thus achieves an optimal trade-off among lightness, accuracy, and deployability, offering a practical paradigm for AI deployment in resource-limited healthcare environments.
0 citationsRead paper