Vision-Language Models (VLMs), exemplified by CLIP, have emerged as foundational for multimodal intelligence. However, their capacity for logical understanding remains significantly underexplored, resulting in critical **''logical blindspots''** that limit their reliability in practical applications. To systematically diagnose this, we introduce **LogicBench**, a comprehensive benchmark with over 50,000 vision-language pairs across 9 logical categories and 4 diverse scenarios: images, videos, anomaly detection, and medical diagnostics. Our evaluation reveals that existing VLMs, even the state-of-the-art ones, fall at over 40 accuracy points below human performance, particularly in challenging tasks like Causality and Conditionality, highlighting their reliance on surface semantics over critical logical structures. To bridge this gap, we propose **LogicCLIP**, a novel training framework designed to boost VLMs' logical sensitivity through advancements in both data generation and optimization objectives. LogicCLIP utilizes logic-aware data generation and a contrastive learning strategy that combines coarse-grained alignment, a fine-grained multiple-choice objective, and a novel logical structure-aware objective. Extensive experiments demonstrate LogicCLIP's substantial improvements in logical comprehension across all LogicBench domains, significantly outperforming baselines. Moreover, LogicCLIP retains, and often surpasses, competitive performance on general vision-language benchmarks, demonstrating that the enhanced logical understanding does not come at the expense of general alignment. We believe LogicBench and LogicCLIP will be important resources for advancing VLM logical capabilities.

The "Fine-Tuning-as-a-Service" paradigm exposes large language models to catastrophic safety degradation from less harmful samples.Alignment-stage defenses address this by proactively injecting adversarial perturbations to bolster the model's inherent robustness against harmful drift.However, existing methods rely on perturbation directions that often conflict with harmful gradients, inadvertently facilitating the acquisition of malicious features rather than suppressing them.To address this issue, we propose Orthogonal and Adaptive Safety Alignment Strategy (OASIS) to mathematically decouple safety enforcement from harmful feature acquisition.By projecting perturbations orthogonal to harmful gradients and concentrating optimization on adaptively selected safetycritical layers, OASIS effectively resolves directional conflicts while maximizing parameter efficiency.Extensive experiments on four LLMs across three datasets (SST2, GSM8K, and AGNews) demonstrate that OASIS reduces the Harmful Score by approximately 60% compared to competitive baselines, while maintaining stable downstream task utility.Our code is publicly available at https://github. com/xiaoroyi/OASIS.

Modeling task-driven attention in driving is a fundamental challenge for both autonomous vehicles and cognitive science. Existing methods primarily predict where drivers look by generating spatial heatmaps, but fail to capture the cognitive motivations behind attention allocation in specific contexts, which limits deeper understanding of attention mechanisms. To bridge this gap, we introduce Explainable Driver Attention Prediction, a novel task paradigm that jointly predicts spatial attention regions (where), parses attended semantics (what), and provides cognitive reasoning for attention allocation (why). To support this, we present W3DA, the first large-scale explainable driver attention dataset. It enriches existing benchmarks with detailed semantic and causal annotations across diverse driving scenarios, including normal conditions, safety-critical situations, and traffic accidents. We further propose LLada, a Large Language model-driven framework for driver attention prediction, which unifies pixel modeling, semantic parsing, and cognitive reasoning within an end-to-end architecture. Extensive experiments demonstrate the effectiveness of LLada, exhibiting robust generalization across datasets and driving conditions. This work serves as a key step toward a deeper understanding of driver attention mechanisms, with significant implications for autonomous driving, intelligent driver training, and human-computer interaction.

Consumer lending service is escalating in E-Commerce platforms due to its capability in enhancing buyers' purchasing power, improving average order value, and increasing revenue of the platforms. Credit risk forecasting and credit limits setting are two fundamental problems in E-Commerce/online consumer lending services. Currently, the majority of institutes rely on two-separate-step methods to resolve. First, build a rating model to evaluate credit risk, and then design heuristic strategies to set credit limits, which requires a large amount of prior knowledge and lacks theoretical justifications. In this paper, we propose an end-to-end multi-view and multi-task learning based approach named MvMoE (Multi-view-aware Mixture-of-Experts network) to solve these two problems simultaneously. First, a multi-view network with a hierarchical attention mechanism is constructed to distill users' heterogeneous financial information into shared hidden representations. Then, we jointly train these two tasks with a view-aware multi-gate mixture-of-experts network and a subsequent progressive network to improve their performances. With the real-world dataset contained 5.44 million users, we investigate the effectiveness of MvMoE. Experimental results exhibit that the proposed model is able to improve AP over 5.60% on credit risk forecasting and MAE over 9.52% on credit limits setting compared with conventional methods. Meanwhile, MvMoE has good interpretability, which better underpins the imperative demands in financial industries.

Financial credit risk assessment serves as the impetus to evaluate the credit admission or potential business failure of customers in order to make early actions prior to the actual financial crisis. It aims to predict the probability that a customer may belong to a high-risk group, which is usually formulated as a binary classification problem. However, due to the lack of high-risk samples, the prevailing models suffer from the severe class-imbalance problem. Oversampling those high-risk users could alleviate this problem but the effect of noise examples is also amplified. In this paper, we propose a novel adversarial data augmentation method to solve the class imbalance problem in financial credit risk assessment. We train a generator for synthetic sample generation with a discriminator to identify real or fake instances. Besides, an auxiliary risk discriminator is trained cooperatively with the generator to assess the credit risk. Experimental results on three real-world datasets demonstrate the effectiveness of the proposed

Credit risk forecasting generally aims to evaluate the default probability of users in financial service. It is usually regarded as a binary classification problem, which suffers from the severe class-imbalance problem due to the extremely limited throngs and the concept drift problem brought by the delayed verification. In this paper, we investigate these problems in credit risk forecasting and propose a semi-supervised meta-learning based approach called TRUST (TRainable Undersampling with Self Training) to resolve. First, it decides whether to sample the data through meta-learning based reinforcement learning. Secondly, it learns the distribution of the data that have not yet shown financial performance via self-training and updates the model trained in the first step. Finally, the updated model is evaluated on the validation dataset, the result of which will be fed back through the evaluator. These three steps will be iterated until the model converges. With the real-world industrial dataset containing 1.75 million users, we investigate the effectiveness of our method. Experimental results exhibit that the proposed method is able to improve AP over 5.94% on credit risk forecasting task compared with the recent methods.

Fraud transactions obtain the rights and interests of e-commerce platforms by illegal ways, and have been the emerging threats to the healthy development of these platforms. Recently, user behavioral data is extensively exploited to detect fraud transactions, and it is usually processed as a sequence consisting of individual actions. However, such sequence-like user behaviors have logical patterns associated with user intentions, which motivates a fine-grained management strategy that binds and cuts off these actions into intention-related segments. In this paper, we devise a tree-like structure named behavior tree to reorganize the user behavioral data, in which a group of successive sequential actions denoting a specific user intention are represented as a branch on the tree. We then propose a novel neural method coined LIC Tree-LSTM(Local Intention Calibrated Tree-LSTM) to utilize the behavior tree for fraud transactions detection. In our LIC Tree-LSTM, the global user intention is captured by an attentional method applied on different branches. Then, we calibrate the entire tree by attentions within tree branches to pinpoint the balance between global and local user intentions. We investigate the effectiveness of LIC Tree-LSTM on a real-world dataset of Alibaba platform, and the experimental results show that our proposed algorithm outperforms state-of-the-art methods in both offline and online modes. Furthermore, our model provides good interpretability which helps us better understand user behaviors.

Default user detection plays one of the backbones in credit risk forecasting and management. It aims at, given a set of corresponding features, e.g., patterns extracted from trading behaviors, predicting the polarity indicating whether a user will fail to make required payments in the future. Recent efforts attempted to incorporate attributed heterogeneous information network (AHIN) for extracting complex interactive features of users and achieved remarkable success on discovering specific default users such as fraud, cash-out users, etc. In this paper, we consider default users, a more general concept in credit risk, and propose a multi-view attributed heterogeneous information network based approach coined MAHINDER to remedy the special challenges. First, multiple views of user behaviors are adopted to learn personal profile due to the endogenous aspect of financial default. Second, local behavioral patterns are specifically modeled since financial default is adversarial and accumulated. With the real datasets contained 1.38 million users on Alibaba platform, we investigate the effectiveness of MAHINDER, and the experimental results exhibit the proposed approach is able to improve AUC over 2.8% and [email protected]=0.1 over 13.1% compared with the state-of-the-art methods. Meanwhile, MAHINDER has as good interpretability as tree-based methods like GBDT, which buoys the deployment in online platforms.

Inquiry performs one of the backbones in current E-commerce websites. Detecting spams in inquiries is essential for these platforms but is surprisingly underexplored by current research. In this paper, we propose a system coined ISTBEL to detect spam inquiries. Motivated by the observations on both behavioral and linguistic differences between spammers and benign users, ISTBEL jointly utilizes the temporal-categorical behavioral sequence and text sequence to accomplish its purpose. At its heart, a variant of LSTM equipped interactive attentions is devised to strengthen the coherence between the categorical and temporal behaviors. A character-level CNN is adopted to capture the capricious linguistic patterns. These two modules are integrated and collaboratively contribute to the final predictions. Our system is trained through an end-to-end manner and can directly utilize raw data as input without tedious feature engineering. We have applied the system to inquiry spam detection on Alibaba.com, and results show ISTBEL is promising in detection spammers and outperforms baselines by a large margin. Meanwhile, ISTBEL can be easily applied to other applications which contain both behavioral and linguistic information.

With the wide usage of microblogging services, microposts grow at high rate and provide a rich source of information related to important social events and trends. However, analyzing microposts is challenging due to its high complexity and large volume. In this paper, we present ImgWordle, an interactive visualization prototype to help people perceive and analyze the image and text information in microblogging services. The prototype extends tag cloud by involving images as well as words, and provides multiple coordinated views for a given event, including keywords of various topics, representative images, geographic sentiments and amounts of microposts, and popular microposts over time. We implement ImgWordle on Sina Weibo, the most popular microblogging service in China, and illustrate the usefulness of this visual interface.

The paper presents a visualization prototype that tackles the issue of public emotion. It emphasizes the exploration process in the aim of creating reflections for the viewer to observe the whole social context as well as individual perspectives. Urban traffic conversation on SNS (Social Networking Site) is our current interest. The prototype we present is made on real post streams captured from Chinese most popular microblog Sina Weibo. After experimenting on two online prototypes we determine a visualizing flow to lead the viewer going through the insights from macro to micro view in three interaction frameworks: City Sentiment, Related Topics, and Post Content. Through showing the visualizing flow with interaction mode, data analysis, and prototype construction, the paper ends with discussing about design considerations in creating reflections on public traffic emotion in visualization prototype.

The Corel Image set is widely used for image annotation performance evaluation although it has been claimed that Corel images are relatively easy to annotate. The aim of this paper is to demonstrate some of the disadvantages of datasets like the Corel set for effective auto-annotation evaluation. We first compare the performance of several annotation algorithms using the Corel set and find that simple near neighbor propagation techniques perform fairly well. A support vector machine (SVM)-based annotation method achieves even better results, almost as good as the best found in the literature. We then build a new image collection using the Yahoo Image Search engine and query-by-single-word searches to create a more challenging annotated set automatically. Then, using three very different image annotation methods, we demonstrate some of the problems of annotation using the Corel set compared with the Yahoo-based training set. In both cases the training sets are used to create a set of annotations for the Corel test set

Traditionally, statistical models for image auto-annotation have been coupled with image segmentation. Considering the performance of the current segmentation algorithms, it can be meaningful to avoid a segmentation stage. In this paper, we propose a new approach to image auto-annotation by building on previously developed statistical models. In this approach, segmentation is avoided through the use of salient regions. The use of the statistical model results in an annotation performance which improves upon our previously proposed saliency-based word propagation technique. We also show that the use of salient regions achieves better results than the use of general image regions or segments