ABSTRACT Automated program repair (APR) aims to automatically detect and fix software defects, thereby improving software reliability and reducing debugging effort. Recently, researchers have explored the retrieval augmentation techniques to enhance large code models' performance in program repair. Existing retrieval augmentation models often inject retrieved information at the input layer, which can lead to input sequence inflation and interfere with the encoder's ability to focus on the core repair task. Meanwhile, learning‐based methods frequently produce unreliable patches, lacking mechanisms to verify or refine low‐confidence outputs during generation. To address these challenges, this paper proposes RADI‐PR, a novel approach that integrates retrieval augmentation and decoding intervention at the model's output layer. RADI‐PR dynamically incorporates relevant repair patterns based on historical fixes and intervenes in low‐confidence generations to enhance both the accuracy and reliability of generated patches. Comprehensive evaluations on four benchmark datasets, including Java, FixJS, Codeflaws and TSSB‐3 M, show that RADI‐PR consistently outperforms baseline methods. RADI‐PR achieves improvements of up to 5.5% in Precision, 3.2% in F1‐score and 2.2% in Accuracy.

The growing use of specialized AI accelerators like Neural Processing Units (NPUs) requires a deep understanding of framework-hardware interactions to achieve optimal performance and energy efficiency. This study investigates these interactions through a comprehensive comparative analysis of TensorFlow, a general-purpose framework, and MindSpore, a hardware-optimized framework, on the Huawei Ascend 910 NPU. We evaluate performance across representative deep learning workloads (ResNet50, YOLOv5, Transformer) at both model and operator levels, analyzing throughput, hardware resource utilization (AI Core, HBM, DDR), and average power consumption. Our findings reveal that MindSpore significantly outperforms TensorFlow in throughput and energy efficiency, primarily driven by its substantially higher utilization of the NPU's High-Bandwidth Memory (HBM). Operator-level analysis confirms that MindSpore's advantage grows with scale, correlating strongly with memory demands. While acknowledging MindSpore's targeted optimizations, this work provides quantitative insights into the critical role of memory hierarchy management for NPU performance and offers practical guidance for optimizing AI systems on specialized hardware accelerators, providing valuable insights for both AI practitioners and the HPC community.

Software vulnerabilities damage the reliability of software systems.Recently, many methods based on deep learning have been proposed for vulnerability detection by learning features from code sequences or various property graphs.However, sequence-based methods frequently fail to consider the structural information of the code, whereas graph-based methods often encounter difficulties in capturing long-distance contextual semantic information.To overcome these limitations, we propose PTCN, a novel vulnerability detection method that integrates Pre-Trained code language model and Convolutional Neural Network (CNN).First, the code tokens sequence and the propagation chains are constructed from the source code in preparation for the subsequent feature extraction.Next, GraphCodeBERT, a graph-based pre-trained code language model, is introduced as a means of learning both the semantic and structural information contained in the code tokens sequence and the propagation chains.Then, a CNN architecture is designed to effectively extract the crucial vulnerability features from the hidden states of the last Transformer layer of GraphCodeBERT for vulnerability detection.The experiments on the public benchmark dataset REVEAL indicate that PTCN outperforms the state-of-the-art methods with respect to the Accuracy and F1-score metrics by 3.41%-15.80%and 22.68%-104.27%,respectively.

Pre-trained code models have achieved promising results in the vulnerability detection field.The prevailing approach is to adapt these models with vulnerability datasets using inefficient full-model fine-tuning.These pre-trained models primarily capture the semantic features of code, while neglecting its structural characteristics.To address these issues, this paper proposes a vulnerability detection method with adapter tuning and enhanced feature learning.First, adapter modules are introduced to UniXcoder and tune only the parameters in the adapters to extract semantic features.This significantly reduces the number of training parameters and better adapts the model to downstream tasks.Then, structural features, including control and data flow information, are extracted from the Program Dependence Graph (PDG) to compensate for the limitations of pre-trained models that rely solely on semantic features.Finally, the semantic and structural features are integrated to train the detection model.The experimental results demonstrate that our proposed method outperforms state-of-the-art approaches and is highly efficient in terms of both training parameters and training data.

The capacity of Deep Learning to automatically learn features from source code has facilitated its extensive utilization in detecting software vulnerabilities.However, existing pre-trained models regard code snippets as token sequences, neglecting the inherent structure of the code.The utilization of graph-based code representations is constrained by the limitations of Graph Neural Networks, particularly the difficulty in capturing long-range dependencies, which renders it challenging to learn grammatical and semantic information from complex code graph representations.This paper proposes a novel multiscale software vulnerability detection method based on supervised contrastive learning.The method integrates local features obtained from code paths with structural features extracted from the Code Property Graph by GraphTrans, thereby achieving a multi-scale feature representation of the source code.Additionally, a supervised contrastive loss function is employed during training in order to fully utilize label information and address the class imbalance problem.The experimental results demonstrate that the proposed method outperforms existing state-of-the-art methods, achieving the highest accuracy, precision, and F1 on the real-world benchmark dataset CodeXGLUE for software vulnerability detection.