Recent advancements in Large Language Models have increasingly demonstrated their potential for event reasoning. However, LLMs still struggle with this task due to inadequate modeling of event structures. Although introducing schema knowledge has been shown to improve event reasoning performance, existing methods rely on predefined schema library, compromising their scalability and lightweight deployment. To address these challenges, we propose SGER, a plug-and-play Schema-Guided Event Reasoning framework. In the schema extraction stage, the model maps event descriptions with diverse surface forms to potential semantic structure representations, achieving an abstract transformation from instances to schemas. The schema prediction stage captures the potential associations between historical event schemas to make forward-looking inferences about possible future event schemas. In the event reasoning stage, we integrate historical events and predicted schemas into prompts to guide LLMs in generating specific, contextually consistent predicted events. Experimental evaluations demonstrate that our framework significantly improves event reasoning performance of LLMs.

Open-world intent discovery is critical for task-oriented dialogue systems, where static intent taxonomies fail to capture emerging user intentions and existing methods show limited generalization beyond predefined label spaces. To address this issue, we propose Unified Generative Intent Discovery (UGID), a unified framework that reformulates intent understanding as a conditional text generation task, enabling both in-domain (IND) intent classification and out-of-domain (OOD) intent discovery within a single architecture. UGID adopts a two-stage training strategy. First, instruction-tuned supervised fine-tuning strengthens semantic discrimination among known intents. Second, a self-play reinforcement learning mechanism simulates iterative user–system interactions to explore, refine, and validate novel intent labels. In addition, a reward design combining semantic fidelity and domain relevance guides the generation process toward coherent and meaningful intent discovery. Experiments on three benchmark datasets demonstrate that UGID consistently achieves strong performance, reaching ACC scores of 79.63%, 82.70% and 87.50% on BANKING, StackOverflow and CLINC, respectively. Compared with the strongest generative baseline, IntentGPT-4, UGID further improves ACC by 14.87 and 4.21 percentage points on BANKING and CLINC, respectively. Ablation studies further verify the effectiveness of the proposed design. Overall, UGID provides an effective and scalable framework for intent understanding in dynamic open-world dialogue scenarios.

Large language models (LLMs) demonstrate significant generative capabilities but often face ethical alignment and robustness challenges. Conventional alignment methods rely on extensive human-annotated data and require retraining, leading to high computational costs and resource demands. Therefore, we propose a novel approach, Decoding Probability Correction (DPC), that aligns frozen LLMs without additional training or annotated data. DPC dynamically adjusts the probability distribution during inference, ensuring the generated content aligns with human values in real-time. Additionally, DPC incorporates a discriminator-based backtracking mechanism, further enhancing content safety by re-evaluating and refining generation choices. Experimental results on datasets such as the HH and AdvBench show that DPC significantly reduces harmful outputs while maintaining high levels of informativeness and helpfulness. The proposed method offers a cost-effective and efficient solution for enhancing the ethical alignment of LLMs in real-world applications.

Federated learning under heterogeneous models, as an innovative approach, aims to break through the constraints of vanilla federated learning on the consistency of model architectures to better accommodate the heterogeneity of data distributions and hardware resource constraints in mobile computing scenarios. While significant attention has been given to backdoor risks in federated learning, the impact on heterogeneous models remains insufficiently investigated, where devices contribute models with varying structures. The reduction in the number of benign local model neurons that the adversary can manipulate through the global model reduces the attack surface. To challenge this issue, we propose a white-box multi-target backdoor attack method, CapsuleBD, against heterogeneous federated learning. Specifically, we design a model decoupling method to separate the benign and malicious task training pipelines through weight reassignment. The model responsible for the benign tasks is structurally larger than the malicious one, resembling a capsule encapsulating harmful substance impacting multiple heterogeneous models. Our comprehensive experiments demonstrate the effectiveness of CapsuleBD in seamlessly embedding triggers into heterogeneous local models, sustaining a remarkable 99.5% average attack success rate against all benign users even with a 50% reduction in the attack space.

Recent studies improve the cross-lingual transfer learning by better aligning the internal representations within the multilingual model or exploring the information of the target language using self-training. However, the alignment-based methods exhibit intrinsic limitations such as non-transferable linguistic elements, while most of the self-training based methods ignore the useful information hidden in the low-confidence samples. To address this issue, we propose CoNLST (Contrastive Negative Learning and Self-Training) to leverage the information of low-confidence samples. Specifically, we extend the negative learning to the metric space by selecting negative pairs based on the complementary labels and then employ self-training to iteratively train the model to converge on the obtained clean pseudo-labels. We evaluate our approach on the widely-adopted cross-lingual benchmark XNLI. The experiment results show that our method improves upon the baseline models and can serve as a beneficial complement to the alignment-based methods.

Hate speech (HS) on social media exacerbates misinformation and baseless prejudices.Evidence-supported counterspeech (CS) is crucial for correcting misinformation and reducing prejudices through facts.Existing methods for generating evidence-supported CS often lack clear guidance with a core claim for organizing evidence and do not adequately address factuality and faithfulness hallucinations in CS within anti-hate contexts.In this paper, to mitigate the aforementioned, we propose F 2 RL, a Factuality and Faithfulness Reinforcement Learning framework for generating claim-guided and evidence-supported CS.Firstly, we generate counter-claims based on hate speech and design a self-evaluation mechanism to select the most appropriate one.Secondly, we propose a coarse-to-fine evidence retrieval method.This method initially generates broad queries to ensure the diversity of evidence, followed by carefully reranking the retrieved evidence to ensure its relevance to the claim.Finally, we design a reinforcement learning method with a triplet-based factuality reward model and a multi-aspect faithfulness reward model.The method rewards the generator to encourage greater factuality, more accurate refutation of HS, consistency with the claim, and better utilization of evidence.Extensive experiments on three benchmark datasets demonstrate that the proposed framework achieves excellent performance in CS generation, with strong factuality and faithfulness.

Zero-shot stance detection aims to determine the stance of previously unseen targets during the inference phase. Achieving effective feature alignment from seen targets to unseen targets is crucial for zero-shot stance detection. In this paper, we propose MSFR, a hierarchical contrastive learning framework, which consists of two core components: inter-aspect contrastive learning for distinguishing aspect-level features and intra-aspect contrastive learning for capturing attribute-level features. Specifically, inter-aspect contrastive learning first maps the global features of an utterance to multiple aspects that influence semantic expression (referred to as aspect-level feature differentiation). This process facilitates the alignment of semantic features across different factors of seen and unseen targets. Intra-aspect contrastive learning enhances the distinguishability of features within the same aspect (referred to as attribute-level feature differentiation) and improves the model’s fine-grained generalization capability. Experimental results demonstrate the superior performance of our model compared to competing baseline models.

In conventional federated learning, each device is restricted to train a network model of a same structure. This greatly hinders the application of federated learning in edge devices and IoT scenarios where the data and devices are quite heterogeneous because of their different hardware equipment and communication networks. At the same time, most of the existing studies about federated learning of heterogeneous models are limited to horizontal heterogeneity which share a highly homogeneous vertical structure. Little work has been done on vertical heterogeneity such as models with different number of functional layers or different connection methods within them, not to mention the integrated heterogeneity scenarios. In DraftFed, a novel draft-based approach is proposed to implement personalized federated learning for integrated heterogeneous models. Unlike traditional federated learning in which the parameters/gradients are exchanged, DraftFed uses drafts as key knowledge to guide mutual learning of models, which makes it suitable for model structure personalization application scenarios..

As one knows, an event often consists of several actions while each action is atomic. Inspired by this insight, we propose a novel framework named Atomic-action-based Contrastive Network model (ACN) for weakly supervised temporal language grounding task to localize the query-related event moment in an untrimmed video, without access to any temporal annotations. Specifically, ACN first determines the accurate moment boundary of each action in a query-agnostic way. This can adequately exploit homogeneous visual cues while impeding the heterogeneity of the query from hurting the atomicity of visual action, i.e., action boundary. To effectively localize the query-related event, we seek the discriminative words in the given query, and explore a composite-grained contrastive module to retrieve those corresponding atomic actions in the common latent space across modalities. This boosts feature discrimination of visual event segment to remove irrelevant action video segments. Experiments on two popular datasets show the efficacy of our model.

Zero-shot stance detection is challenging because it requires detecting the stance of previously unseen targets in the inference phase. The ability to learn transferable target-invariant features is critical for zero-shot stance detection. In this paper, we propose a stance detection approach that can efficiently adapt to unseen targets, the core of which is to capture target-invariant syntactic expression patterns as transferable knowledge. Specifically, we first augment the data by masking the topic words of sentences, and then feed the augmented data to an unsupervised contrastive learning module to capture transferable features. Besides, to fit a specific target, we encode the raw text as target-specific features. Finally, we adopt an attention mechanism, which combines syntactic expression patterns with target-specific features to obtain enhanced features for predicting previously unseen targets. Experiments demonstrate that our model outperforms competitive baselines on four benchmark datasets.

Hate speech detection is hampered by the scarcity and topical and lexical biases of annotated data, leading to poor generalization. It is imperative to devise a cross-domain approach to solve this problem. The ability to learn transferable knowledge is critical for cross-domain hate speech detection. In this work, We propose a domain-adaptive dependency graph method based on post-hoc explanation (DPDG). We extract post-hoc explanations from fine-tuned BERT classifiers as the importance score for hate representation. Based on these, we construct in-domain graph and cross-domain graph to better learn in-domain hate representation and adapt to the target domain respectively. Finally, we use interactive GCN blocks to interactively and adaptively learn and adjust the domain adaptive graph representation. The results of cross-domain experiments on multiple domains show that our proposed model outperforms competitive baselines in cross-domain hate speech detection.

State and local governments have imposed health policies to contain the spread of COVID-19 since it had a serious impact on human daily life. However, the public stance on these measures may be time-varying. It is likely to escalate the infection in the area where the public is negative or resistant. To take advantage of the correlation between public stance on health policies and the COVID-19 statistics, we propose a novel framework, Multitask Learning Neural Networks for Pandemic Prediction with Public Stance Enhancement (MP3), which is composed of three modules: (1) Stance awareness module to make stance detection on health policies from users' tweets in social media and convert them into a stance time series. (2) Temporal feature extraction module that applies Convolution Neural Network and Recurrent Neural Network to extract and fuse local patterns and long-term correlations from COVID-19 statistics. Moreover, a Stance Latency-aware Attention is proposed to capture dynamic social effects and fuse them with temporal features. (3) Multi-task prediction module to adopt Graph Convolution Network to model the spread of pandemic and employ multi-task learning to simultaneously predict COVID-19 statistics and the trend of public stance on health policies. The proposed framework outperforms state-of-the-art baselines on both confirmed cases and deaths prediction tasks.

Zero-shot stance detection (ZSSD) is an important research problem that requires algorithms to have good stance detection capability even for unseen targets. In general, stance features can be grouped into two types: target-invariant and target-specific. Target-invariant features express the same stance regardless of the targets they are associated with, and such features are general and transferable. On the contrary, target-specific features will only be directly associated with specific targets. Therefore, it is crucial to effectively mine target-invariant features in texts in ZSSD. In this paper, we develop a method based on contrastive learning to mine certain transferable target-invariant expression features in texts from two dimensions of sentiment and stance and then generalize them to unseen targets. Specifically, we first grouped all texts into several types in terms of two orthogonal dimensions: sentiment polarity and stance polarity. Then we devise a supervised contrastive learning-based strategy to capture each type's common and transferable expressive features. Finally, we fuse the above-mentioned expressive features with the semantic features of the original texts about specific targets to deal with the stance detection for unseen targets. Extensive experiments on three benchmark datasets show that our proposed model achieves the state-of-the-art performance on most datasets. Code and other resources are available on GitHub11https://github.com/zoujiaying1995/sscl-project.

The accurate forecasting of infectious epidemic diseases is the key to effective control of the epidemic situation in a region.Most existing methods ignore potential dynamic dependencies between regions or the importance of temporal dependencies and inter-dependencies between regions for prediction.In this paper, we propose an Interand Intra-Series Embeddings Fusion Network (SEFNet) to improve epidemic prediction performance.SEFNet consists of two parallel modules, named Inter-Series Embedding Module and Intra-Series Embedding Module.In Inter-Series Embedding Module, a multiscale unified convolution component called Region-Aware Convolution is proposed, which cooperates with self-attention to capture dynamic dependencies between time series obtained from multiple regions.The Intra-Series Embedding Module uses Long Short-Term Memory to capture temporal relationships within each time series.Subsequently, we learn the influence degree of two embeddings and fuse them with the parametric-matrix fusion method.To further improve the robustness, SEFNet also integrates a traditional autoregressive component in parallel with nonlinear neural networks.Experiments on four real-world epidemic-related datasets show SEFNet is effective and outperforms state-of-the-art baselines.

AR navigation has been widely used in mobile devices. However, users sometimes immerse in the navigation interface and ignore probable dangers in the real environment. It is necessary to remind users of potential dangers and avoid accidents. Most of existing works focus on how to effectively guide users in AR but few concern about design of danger reminder. In this paper, we build a virtual AR navigation system and compare user experience on different types of obstacle reminder. Furthermore, we compare the influence of color, motion and appearance distance on effectiveness of the reminder. Results show that red color and bi-color are more obvious than blue color for reminder. Motion such as flickering effect helps enhance remind effectiveness.