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.

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.

Students come to science classrooms with ideas informed by their prior instruction and everyday observations. Following constructivist pedagogy, assessments that encourage students to elaborate their ideas, distinguish among them, and link the most promising ones can capture students' potential and help teachers plan their lessons. In this investigation, we study an assessment that engages students in a dialog to refine their response to a Knowledge Integration (KI) question. Our Research Practice Partnership (RPP) initially trained a Natural Language Processing (NLP) idea detection model on 1218 student responses from 5 schools and identified 13 student ideas. The original model had an overall micro-averaged F-score of 0.7634. After classroom testing, three RPP expert teachers with 10+ years of experience reviewed the classroom data and expanded the model, adding six additional ideas including two that they described as precursor ideas because they foreshadowed more sophisticated reasoning. We trained the idea detection model on these 19 ideas using a dataset from 13 teachers and 1206 students across 8 public schools. The updated model had a somewhat lower overall micro-averaged F-score of 0.7297. The two precursor ideas were among the top four detected ideas. The assessment, using the updated model, guided students to express significantly more ideas. A regression model showed that the updated model was associated with greater KI score gains. Expanding the model, thus, created an assessment that motivated students to express more ideas and to achieve higher KI scores. It also provides teachers with deeper insights into their students' understanding of science.

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..

Abstract In conventional federated learning, each device is restricted to train a network model of the same structure. This greatly hinders the application of federated learning where the data and devices are quite heterogeneous because of their different hardware equipment and communication networks. At the same time, existing studies have shown that transmitting all of the model parameters not only has heavy communication costs, but also increases risk of privacy leakage. We propose a general framework for personalized federated learning (PerHeFed), which enables the devices to design their local model structures autonomously and share sub-models without structural restrictions. In PerHeFed, a simple-but-effective mapping relation and a novel personalized sub-model aggregation method are proposed for heterogeneous sub-models to be aggregated. By dividing the aggregations into two primitive types (i.e., inter-layer and intra-layer), PerHeFed is applicable to any combination of heterogeneous convolutional neural networks, and we believe that this can satisfy the personalized requirements of heterogeneous models. Experiments show that, compared to the state-of-the-art method (e.g., FLOP), in non-IID data sets our method compress ≈ 50% of the shared sub-model parameters with only a 4.38% drop in accuracy on SVHN dataset and on CIFAR-10, PerHeFed even achieves a 0.3% improvement in accuracy. To the best of our knowledge, our work is the first general personalized federated learning framework for heterogeneous convolutional networks, even cross different networks, addressing model structure unity in conventional federated learning.

The computing power of various Internet of Things (IoT) devices is quite different. To enable IoT devices with lower computing power to perform machine learning, all nodes can only train smaller models, which results in the waste of computing power for high-performance devices. In this article, a heterogeneous model fusion federated learning (HFL) mechanism is proposed. Each node trains learning models of different scales according to its own computing capabilities. After receiving the gradient trained by each node, the parameter server (PS) corrects the received gradient with the repeat matrix, and then update the corresponding region of the global model according to the mapping matrix. After all update operations are over, the PS assigns the compressed model to the corresponding node. This article uses a variety of experimental schemes to evaluate the proposed method, including three data sets, two model structures, and three computational complexity levels. The proposed method have been proved it not only maximizes the use of the unbalanced computing power of edge nodes, but also enables different structural models to compensate for the shortcomings of others, improving the overall performance.