Visual hallucination remains a major obstacle to the reliability of Large Vision-Language Models (LVLMs). We argue that this issue originates from a fundamental statistical misspecification: the conventional softmax attention implicitly assumes i.i.d. noise, yet real LVLM attention patterns exhibit structured and competitive biases (e.g., attention sinks) that violate this assumption. To address this mismatch, we introduce Latent Attention Denoising (LAD), a principled and training-free framework that recasts attention calibration as a one-step score-based denoising process. LAD employs an interpretable energy function to derive an analytic score and applies a single Langevin-inspired update to actively steer corrupted attention logits toward more faithful configurations. This intervention imposes negligible computational overhead and operates at a speed comparable to standard greedy decoding. Extensive evaluations across diverse architectures confirm that LAD achieves superior performance on both generative and discriminative tasks, effectively mitigating hallucinations while maintaining efficiency comparable to standard decoding.

Tokenized graph Transformers have advanced node classification by transforming graphs into token sequences, but existing methods suffer from limited flexibility due to single-type token generation, which captures partial graph information and requires tailored modifications. To address this, we propose NTFormer, a novel graph Transformer with a dedicated token generator called Node2Par. Node2Par constructs diverse token sequences for each node using multiple token elements (i.e., neighborhood tokens and node tokens) from both topology view and attribute view, enabling comprehensive expression of graph features from multi-perspectives. Leveraging the outputs of Node2Pars, NTFormer adopts a standard Transformer backbone without additional graph-aware modules and a learnable information fusion strategy to adaptively learn expressive node representations from generated different token sequences, eliminating the need for tailored encoding strategies. Extensive experiments on benchmark datasets including homophily and heterophily graphs showcase that NTFormer outperforms representative graph Transformers and GNNs in node classification.

Along with the sustained occurrence of black swan events in the decentralized application ecosystem, smart contract security is a growing concern. Traditional solutions mainly rely on predefined rules, while highly accurate, require intensive manual code analysis. Machine learning methods (mainly based on BERT) leverage semantic and contextual information, but overlook crucial code structure features, which are critical for identifying vulnerability. Furthermore, the presence of useless or harmful attention heads in the BERT model leads to less robust predictions and slows down processing speeds. We propose a novel method named Pruning attention Heads based on Semantic and Code Structure (PHSCS) for Smart Contract Vulnerability Detection. Specifically, we introduce a new structure-aware pre-training programming language task, Variable Edge Prediction, which bypasses the use of data flow nodes as input and directly predicts data flow edges between variables, aiming to efficiently learn code structure while ensuring the ability to process extensive code. Additionally, we present a pruning strategy to optimize BERT, tailored to the semantic and structural peculiarities of code. By employing Taylor Expansion for evaluating attention heads' significance and guiding their pruning, iteratively refined the BERT model. Experiment results on 8 vulnerability types illustrate that the PHSCS method surpasses state-of-the-art methods.

Smart contracts are crucial for managing sensitive financial transactions. However, these contracts are inherently vulnerable which can cause significant security risks. Traditional methods for detecting smart contract vulnerabilities depend on expert-defined rules, which are often complicated and limited by human experts' individual experience. In contrast, deep learning-based approaches automatically extract intricate feature representations, greatly improving detection efficiency and accuracy. Nevertheless, these approaches rely on access to large amounts of high-quality labeled data. They are less effective when facing new types of vulnerabilities where labeled data are scarce. To this end, we propose the Semi-supervised Tuning (SST) approach for smart contract vulnerability detection. It first leverages a source model trained on labeled source data to extract features for new vulnerabilities. Subsequently, it performs semi-supervised learning to explore the feature structure of unlabeled data. In particular, SST groups contract code features and constructs a shared feature queue containing labeled and unlabeled contracts to explore the complete feature structure and guide model training. Extensive experimental evaluations based on two real-world datasets containing eleven smart contract vulnerabilities demonstrate that SST is significantly more advantageous compared to eight state-of-the-art baseline methods, outperforming them by 24.12% in terms of F1 scores.

Existing methods for 3D human mesh recovery from video rely mainly on Recurrent Neural Networks (RNNs) or Transformers. However, due to the limitations of RNNs in temporal modeling and the dense strategies of traditional attention mechanisms, these methods struggle to efficiently model human motion in videos. To address this issue, we propose a novel method that exploits sparse deformable attention mechanisms, efficiently extracting critical spatio-temporal mesh information from the input videos. Specifically, our method consists of two novel modules: the 3D Deformable Mesh Attention (3D-DMA) module and the 2D Deformable Mesh Attention (2D-DMA) module. The 3D-DMA module adaptively extracts human mesh information by sparsely aggregating the features at varied spatio-temporal locations with attentions, while the 2D-DMA module captures the mesh contexts with the attentions of features at sparsely sampled spatial locations. By fusing the outputs of the two modules, we obtain accurate and smooth human body estimations. Extensive experiments show that our model outperforms previous state-of-the-art methods on three widely used benchmarks.

Semi-Supervised Video Paragraph Grounding (SSVPG) aims to localize multiple sentences in a paragraph from an untrimmed video with limited temporal annotations. Existing methods focus on teacher-student consistency learning and video-level contrastive loss, but they overlook the importance of perturbing query contexts to generate strong supervisory signals. In this work, we propose a novel Context Consistency Learning (CCL) framework that unifies the paradigms of consistency regularization and pseudo-labeling to enhance semi-supervised learning. Specifically, we first conduct teacher-student learning where the student model takes as inputs strongly-augmented samples with sentences removed and is enforced to learn from the adequately strong supervisory signals from the teacher model. Afterward, we conduct model retraining based on the generated pseudo labels, where the mutual agreement between the original and augmented views’ predictions is utilized as the label confidence. Extensive experiments show that CCL outperforms existing methods by a large margin.

As a complex neural network system, the brain regions and genes collaborate to effectively store and transmit information. We abstract the collaboration correlations as the brain region gene community network (BG-CN) and present a new deep learning approach, such as the community graph convolutional neural network (Com-GCN), for investigating the transmission of information within and between communities. The results can be used for diagnosing and extracting causal factors for Alzheimer's disease (AD). First, an affinity aggregation model for BG-CN is developed to describe intercommunity and intracommunity information transmission. Second, we design the Com-GCN architecture with intercommunity convolution and intracommunity convolution operations based on the affinity aggregation model. Through sufficient experimental validation on the AD neuroimaging initiative (ADNI) dataset, the design of Com-GCN matches the physiological mechanism better and improves the interpretability and classification performance. Furthermore, Com-GCN can identify lesioned brain regions and disease-causing genes, which may assist precision medicine and drug design in AD and serve as a valuable reference for other neurological disorders.

With the widespread application of smart contracts and the expansion of asset management scale, various new attacks continue to emerge, and a method that can adapt to new vulnerabilities more quickly is urgently needed. Although deep learning methods have shown superior performance in vulnerability detection, their dependence on a large number of labeled samples limits their applicability in new vulnerability scenarios. Therefore, this article introduces a dual adversarial domain adaptation (DADA) approach. This approach consists of two generators and two discriminators. First, the source generator is pretrained with extensive labeled known vulnerability samples in the source domain to extract discriminative features, and its parameters are shared with the target generator. Subsequently, the features of the source and target domain samples are input into the source discriminator, and the target generator is guided to learn domain-invariant features through adversarial training; at the same time, the target domain discriminator is introduced to further weaken its dependence on the distribution of source domain features, thereby improving its adaptability to new vulnerabilities. We conducted experimental evaluations on public datasets, and the results show that our proposed method outperforms six mainstream deep learning-based detection methods. We applied domain adaptation methods to smart contract vulnerability detection for the first time, providing a reference for small sample learning in this field in the future.

Cross‐project defect prediction (CPDP) is an attractive research area in software testing. It identifies defects in projects with limited labeled data (target projects) by utilizing predictive models from data‐rich projects (source projects). Existing CPDP methods based on transfer learning mainly rely on the assumption of a unimodal distribution and consider the case where the feature distribution has one obvious peak. However, in actual situations, the feature distribution of project samples often exhibits multiple peaks that cannot be ignored. It manifests as a multimodal distribution, making it challenging to align distributions between different projects. To address this issue, we propose a balanced adversarial tight‐matching model for CPDP. Specifically, this method employs multilinear conditioning to obtain the cross‐covariance of both features and classifier predictions, capturing the multimodal distribution of the feature. When reducing the captured multimodal distribution differences, pseudo‐labels are needed, but pseudo‐labels have uncertainty. Therefore, we additionally add an auxiliary classifier and attempt to generate pseudo‐labels using a pseudo‐label strategy with less uncertainty. Finally, the feature generator and two classifiers undergo adversarial training to align the multimodal distributions of different projects. This method outperforms the state‐of‐the‐art CPDP model used on the benchmark dataset.

Software defect prediction (SDP) is a crucial phase preceding the launch of software products. Cross‐project defect prediction (CPDP) is introduced for the anticipation of defects in novel projects lacking defect labels. CPDP can use defect information of mature projects to speed up defect prediction for new projects. So that developers can quickly get the defect information of the new project, so that they can test the software project pertinently. At present, the predominant approaches in CPDP rely on deep learning, and the performance of the ultimate model is notably affected by the quality of the training dataset. However, the dataset of CPDP not only has few samples but also has almost no label information in new projects, which makes the general deep‐learning‐based CPDP model not ideal. In addition, most of the current CPDP models do not fully consider the enrichment of classification boundary samples after cross‐domain, leading to suboptimal predictive capabilities of the model. To overcome these obstacles, we present contrastive learning pretraining for CPDP (ConCPDP), a CPDP method integrating contrastive pretraining and category boundary adjustment. We first perform data augmentation on the source and target domain code files and then extract the enhanced data as an abstract syntax tree (AST). The AST is then transformed into an integer sequence using specific mapping rules, serving as input for the subsequent neural network. A neural network based on bidirectional long short‐term memory (Bi‐LSTM) will receive an integer sequence and output a feature vector. Then, the feature vectors are input into the contrastive module to optimise the feature extraction network. The pretrained feature extractor can be fine‐tuned by the maximum mean discrepancy (MMD) between the feature distribution of the source domain and the target domain and the binary classification loss on the source domain. This paper conducts a large number of experiments on the PROMISE dataset, which is commonly used for CPDP, to validate ConCPDP’s efficacy, achieving superior results in terms of F 1 measure, area under curve (AUC), and Matthew’s correlation coefficient (MCC).

As mobile applications evolve rapidly, their fast iterative update nature leads to an increase in software defects. Just-In-Time Software Defect Prediction (JIT-SDP) offers immediate feedback on code changes. For new applications without historical data, researchers have proposed Cross-Project JIT-SDP (CP JIT-SDP). Existing CP JIT-SDP approaches are designed for offline scenarios where target data is available in advance. However, target data in real-world applications usually arrives online in a streaming manner, making online CP JIT-SDP face cross-project distribution differences and target project data concept drift challenges in online scenarios. These challenges often co-exist during application development, and their interactions cause model performance to degrade. To address these issues, we propose an online CP JIT-SDP framework called COTL. Specifically, COTL consists of two stages: offline and online. In the offline stage, the cross-domain structure preserving projection algorithm is used to reduce the cross-project distribution differences. In the online stage, target data arrives sequentially over time. By reducing the differences in marginal and conditional distributions between offline and online data for target project, concept drift is mitigated and classifier weights are updated online. Experimental results on 15 mobile application benchmark datasets show that COTL outperforms 13 benchmark methods on four performance metrics.

Abstract Cross‐project defect prediction (CPDP) is an important research direction in software defect prediction. Traditional CPDP methods based on hand‐crafted features ignore the semantic information in the source code. Existing CPDP methods based on the deep learning model may not fully consider the differences among projects. Additionally, these methods may not accurately classify the samples near the classification boundary. To solve these problems, the authors propose a model based on multi‐adaptation and nuclear norm (MANN) to deal with samples in projects. The feature of samples were embedded into the multi‐core Hilbert space for distribution and the multi‐kernel maximum mean discrepancy method was utilised to reduce differences among projects. More importantly, the nuclear norm module was constructed, which improved the discriminability and diversity of the target sample by calculating and maximizing the nuclear norm of the target sample in the process of domain adaptation, thus improving the performance of MANN. Finally, extensive experiments were conducted on 11 sizeable open‐source projects. The results indicate that the proposed method exceeds the state of the art under the widely used metrics.

With the rapid development of real-time bidding (RTB) in online advertising, learning the distribution of market price has attracted wide attention, since it plays a critical role in designing bidding strategies. One important problem is the right-censored issue in which the true market price can only be observed by the winner of the auction. To address this, existing studies often use Kaplan–Meier estimation (KM), which is one of the best options for survival analysis. However, these approaches depend on counting sample segments and cannot provide accurate predictions for each individual bid request. To enhance the prediction ability, we propose an original method to build the KM for each bid request by predicting (1) the probability of winning an auction at a specific market price, and (2) the probability of losing an auction at a certain bid price. To deal with the high-dimensional sample data common in RTB scenarios, we design a Markov network to calculate these two probabilities. Extensive experiments on two public datasets demonstrate that the proposed approach significantly outperforms state-of-the-art baselines in terms of various metrics, including Wasserstein distance, KL-divergence, average negative log probability and mean squared error.

Expensive optimization problems arise in diverse fields, and the expensive computation in terms of function evaluation poses a serious challenge to global optimization algorithms. In this article, a simple yet effective optimization algorithm for computationally expensive optimization problems is proposed, which is called the neighborhood regression optimization algorithm. For a minimization problem, the proposed algorithm incorporates the regression technique based on a neighborhood structure to predict a descent direction. The descent direction is then adopted to generate new potential offspring around the best solution obtained so far. The proposed algorithm is compared with 12 popular algorithms on two benchmark suites with up to 30 decision variables. Empirical results demonstrate that the proposed algorithm shows clear advantages when dealing with unimodal and smooth problems, and is better than or competitive with other peer algorithms in terms of the overall performance. In addition, the proposed algorithm is efficient and keeps a good tradeoff between solution quality and running time.

The rapid rise of real-time bidding-based online advertising has brought significant economic benefits and attracted extensive research attention. From the perspective of an advertiser, it is crucial to perform accurate utility estimation and cost estimation for each individual auction in order to achieve cost-effective advertising. These problems are known as the click through rate (CTR) prediction task and the market price modeling task, respectively. However, existing approaches treat CTR prediction and market price modeling as two independent tasks to be optimized without regard to each other, thus resulting in suboptimal performance. Moreover, they do not make full use of unlabeled data from the losing bids during estimations, which makes them suffer from the sample selection bias issue. To address these limitations, we propose Multi-task Advertising Estimator (MTAE), an end-to-end joint optimization framework which performs both CTR prediction and market price modeling simultaneously. Through multi-task learning, both estimation tasks can take advantage of knowledge transfer to achieve improved feature representation and generalization abilities. In addition, we leverage the abundant bid price signals in the full-volume bid request data and introduce an auxiliary task of predicting the winning probability into the framework for unbiased learning. Through extensive experiments on two large-scale real-world public datasets, we demonstrate that our proposed approach has achieved significant improvements over the state-of-the-art models under various performance metrics.

In this paper, we tackle the task of natural language video localization (NLVL): given an untrimmed video and a description language query, the goal is to localize the temporal segment within the video that best describes the natural language description. NLVL is challenging at the intersection of language and video understanding because a video may contain multiple segments of interests and the language may describe complicated temporal dependencies. Though existing approaches have achieved good performance, most of them did not fully consider the inherent differences between language and video modalities. Here, we propose Moment Relation Network (MRN) to reduce the divergence of the probability distribution of these two modalities. Specifically, MRN trains video and language subnets, and then uses transfer learning techniques to map the extracted features into an embedding-shared space where we calculate the similarity of two modalities using Mahalanobis distance metric, which is used to localize moments. Extensive experiments on benchmark datasets show that the proposed MRN significantly outperforms the state-of-the-art under the widely used metrics by a large margin.

Evolution strategies have been demonstrated to have the strong ability to roughly train deep neural networks and well accomplish reinforcement learning tasks. However, existing evolution strategies designed specially for deep reinforcement learning only involve the plain variants which can not realize the adaptation of mutation strength or other advanced techniques. The research of applying advanced and effective evolution strategies to reinforcement learning in an efficient way is still a gap. To this end, this paper proposes a restart-based rank-1 evolution strategy for reinforcement learning. When training the neural network, it adapts the mutation strength and updates the principal search direction in a way similar to the momentum method, which is an ameliorated version of stochastic gradient ascent. Besides, two mechanisms, i.e., the adaptation of the number of elitists and the restart procedure, are integrated to deal with the issue of local optima. Experimental results on classic control problems and Atari games show that the proposed algorithm is superior to or competitive with state-of-the-art algorithms for reinforcement learning, demonstrating the effectiveness of the proposed algorithm.

Although the machine learning-based software defect prediction (SDP) method has shown promising value in software engineering, yet challenges remain.To improve the performance of SDP, some researchers have used deep learning algorithms to extract the semantic and structural features of the program.However, in more practical cross-project defect prediction (CPDP) tasks, whether deep learning-generated features can be directly used should be explored due to the data distribution shift that usually exists in different projects.In this paper, we propose a Transferable Hybrid Features Learning with Convolutional Neural Network (CNN-THFL) framework to conduct CPDP.Specially, CNN-THFL mines deep learning-generated features from token vectors extracted from programs' abstract syntax trees via convolutional neural network.Furthermore, CNN-THFL learns the transferable joint features simultaneously considering deep learning-generated and handcrafted features by applying a transfer component analysis algorithm.Finally, the features generated by CNN-THFL are fed to the classifier to train a defect prediction model.Extensive experiments verify that CNN-THFL can outperform referential methods on 72 pairs of CPDP tasks formed by 9 open-source projects.

Using classification methods to predict software defect is receiving a great deal of attention and most of the existing studies primarily conduct prediction under the within‐project setting. However, there usually had no or very limited labelled data to train an effective prediction model at an early phase of the software lifecycle. Thus, cross‐project defect prediction (CPDP) is proposed as an alternative solution, which is learning a defect predictor for a target project by using labelled data from a source project. Differing from previous CPDP methods that mainly apply instances selection and classifiers adjustment to improve the performance, in this study, the authors put forward a novel distribution–adaptation‐based CPDP approach, joint distribution matching (JDM). Specifically, JDM aims to minimise the joint distribution divergence between the source and target project to improve the CPDP performance. By constructing an adaptive weight vector for the instances of the source project, JDM can be effective and robust at reducing marginal distribution discrepancy and conditional distribution discrepancy simultaneously. Extensive experiments verify that JDM can outperform related distribution–adaptation‐based methods on 15 open‐source projects that are derived from two types of repositories.

Machine-learning-based software defect prediction (SDP) methods are receiving great attention from the researchers of intelligent software engineering. Most existing SDP methods are performed under a within-project setting. However, there usually is little to no within-project training data to learn an available supervised prediction model for a new SDP task. Therefore, cross-project defect prediction (CPDP), which uses labeled data of source projects to learn a defect predictor for a target project, was proposed as a practical SDP solution. In real CPDP tasks, the class imbalance problem is ubiquitous and has a great impact on performance of the CPDP models. Unlike previous studies that focus on subsampling and individual methods, this study investigated 15 imbalanced learning methods for CPDP tasks, especially for assessing the effectiveness of imbalanced ensemble learning (IEL) methods. We evaluated the 15 methods by extensive experiments on 31 open-source projects derived from five datasets. Through analyzing a total of 37504 results, we found that in most cases, the IEL method that combined under-sampling and bagging approaches will be more effective than the other investigated methods.

In this paper, we present a multimodal method based on densely connected convolution and bidirectional long-short-term-memory (BLSTM) for gesture recognition. The proposed method learns spatial features of gestures through the densely connected convolutional network, and then learns long-term temporal features by BLSTM network. In addition, fusion methods are evaluated on our model, and we find that fusion of features with different information can significantly improve the recognition accuracy. This purely data driven approach achieves state-of-the-art recognition accuracy on the ChaLearn LAP 2014 dataset and the Sheffield Kinect gesture (SKIG) dataset. (98.80% on the ChaLearn LAP and 99.07% on SKIG).

Software defect prediction (SDP) technology is receiving widely attention and most of SDP models are trained on data from the same project. However, at an early phase of the software lifecycle, there are little to no within‐project training data to learn an available supervised defect‐prediction model. Thus, cross‐project defect prediction (CPDP), which is learning a defect predictor for a target project by using labelled data from a source project, has shown promising value in SDP. To better perform the CPDP, most current studies focus on filtering instances or selecting features to weaken the impact of irrelevant cross‐project data. Instead, the authors propose a novel multiple‐components weights (MCWs) learning model to analyse the varying auxiliary power of multiple components in a source project to construct a more precise ensemble classifiers for a target project. By combining the MCW model with kernel mean matching algorithm, their proposed approach adjusts the source‐instance weights and source‐component weights to jointly alleviate the negative impacts of irrelevant cross‐project data. They conducted comprehensive experiments by employing 15 real‐world datasets to demonstrate the advantages and effectiveness of their proposed approach.

This demo presents WiSee, a novel human-computer interaction system that leverages wireless networks (e.g., Wi-Fi), to enable sensing and recognition of human gestures and motion. Since wire- less signals do not require line-of-sight and can traverse through walls, WiSee enables novel human-computer interfaces for remote device control and building automation. Further, it achieves this goal without requiring instrumentation of the human body with sensing devices. We integrate WiSee with applications and demonstrate how WiSee enables users to use gestures and control applications including music players and gaming systems. Specifically, our demo will allow SIGCOMM attendees to control a music player and a lighting control device using gestures.