Selected work

Representative Papers

Revisiting Vul-RAG: Reproducibility and Replicability of RAG-based Vulnerability Detection with Open-Weight Models

Jun 03, 2026

This study addresses the limited reproducibility and generalizability of existing large language model (LLM)-based vulnerability detection approaches, which often rely on closed-source models and proprietary APIs. The authors systematically reproduce the Vul-RAG framework in a fully local, open-weight setting and, for the first time, evaluate its feasibility across a diverse set of open-source LLMs—including code-specific, general-purpose, and reasoning-oriented models—using a standardized evaluation protocol. Experimental results reveal that all models converge to a performance plateau around 0.30 pairwise accuracy, indicating a saturation effect between model scale and vulnerability detection efficacy. This challenges the prevailing “bigger is better” assumption and suggests that merely increasing model capacity yields diminishing returns in improving detection performance.

0 citationsRead paper

SFG-ROS: A Resource-Aware Framework for Dense Multi-Agent Perception

May 22, 2026

Standard ROS 2 faces challenges such as network congestion, naming conflicts, and high computational overhead in dense multi-agent perception scenarios. This work proposes a resource-aware collaborative perception framework that employs structured fully qualified names to achieve traffic isolation, integrates Fast DDS directional routing with lightweight inter-process communication (IPC) optimizations, and introduces on-demand centralized decoding alongside hardware abstraction containers to enable zero-configuration deployment across heterogeneous accelerators. The proposed approach reduces network traffic control complexity to O(1), decreases per-subscriber CPU overhead by 72.3% compared to standard ROS 2, and maintains low latency, thereby significantly enhancing system scalability and efficiency.

0 citationsRead paper

StixelNExT++: Lightweight Monocular Scene Segmentation and Representation for Collective Perception

Jul 09, 2025

To address the high computational cost and low representation efficiency of monocular 3D scene understanding in autonomous driving collective perception, this paper proposes a lightweight monocular 3D scene representation method. Our approach integrates fine-grained 3D Stixel units with a learnable clustering mechanism, enabling semantic-aware adaptive clustering that compresses scene representations while improving object segmentation accuracy. We design a lightweight neural network that takes a single RGB image as input and jointly leverages depth estimation and LiDAR-based self-supervised ground truth to efficiently generate Stixel representations—natively supporting multimodal outputs including point clouds and bird’s-eye-view (BEV) maps. Evaluated on the Waymo Open Dataset within a 30-meter range, our method achieves state-of-the-art performance with only 10 ms inference time per frame, striking an optimal balance among real-time efficiency, accuracy, and compatibility with collaborative perception systems.

0 citationsRead paper

Recent publications

Latest Papers

Revisiting Vul-RAG: Reproducibility and Replicability of RAG-based Vulnerability Detection with Open-Weight Models

Jun 03, 2026

This study addresses the limited reproducibility and generalizability of existing large language model (LLM)-based vulnerability detection approaches, which often rely on closed-source models and proprietary APIs. The authors systematically reproduce the Vul-RAG framework in a fully local, open-weight setting and, for the first time, evaluate its feasibility across a diverse set of open-source LLMs—including code-specific, general-purpose, and reasoning-oriented models—using a standardized evaluation protocol. Experimental results reveal that all models converge to a performance plateau around 0.30 pairwise accuracy, indicating a saturation effect between model scale and vulnerability detection efficacy. This challenges the prevailing “bigger is better” assumption and suggests that merely increasing model capacity yields diminishing returns in improving detection performance.

0 citationsRead paper

SFG-ROS: A Resource-Aware Framework for Dense Multi-Agent Perception

May 22, 2026

Standard ROS 2 faces challenges such as network congestion, naming conflicts, and high computational overhead in dense multi-agent perception scenarios. This work proposes a resource-aware collaborative perception framework that employs structured fully qualified names to achieve traffic isolation, integrates Fast DDS directional routing with lightweight inter-process communication (IPC) optimizations, and introduces on-demand centralized decoding alongside hardware abstraction containers to enable zero-configuration deployment across heterogeneous accelerators. The proposed approach reduces network traffic control complexity to O(1), decreases per-subscriber CPU overhead by 72.3% compared to standard ROS 2, and maintains low latency, thereby significantly enhancing system scalability and efficiency.

0 citationsRead paper

StixelNExT++: Lightweight Monocular Scene Segmentation and Representation for Collective Perception

Jul 09, 2025

To address the high computational cost and low representation efficiency of monocular 3D scene understanding in autonomous driving collective perception, this paper proposes a lightweight monocular 3D scene representation method. Our approach integrates fine-grained 3D Stixel units with a learnable clustering mechanism, enabling semantic-aware adaptive clustering that compresses scene representations while improving object segmentation accuracy. We design a lightweight neural network that takes a single RGB image as input and jointly leverages depth estimation and LiDAR-based self-supervised ground truth to efficiently generate Stixel representations—natively supporting multimodal outputs including point clouds and bird’s-eye-view (BEV) maps. Evaluated on the Waymo Open Dataset within a 30-meter range, our method achieves state-of-the-art performance with only 10 ms inference time per frame, striking an optimal balance among real-time efficiency, accuracy, and compatibility with collaborative perception systems.

0 citationsRead paper