Referring multimodal large language models enable users to ground queries to specific image regions via spatial prompts, supporting fine-grained referring dialogue. However, existing methods rely on extensive fine-tuning to mitigate attention distraction, which incurs high computational costs and limits adaptability. Without sufficient training data, irrelevant regions in single images easily divert model focus, leading to redundant outputs or hallucinations. To address this, we propose CoreGaze, a training-free framework that simulates human visual gaze diffusion for fine-grained comprehension. First, CoreGaze constructs a sparse semantic graph from visual tokens, modeling region-wise affinities via thresholded similarity. It then maps the user’s visual prompt to a core subgraph with amplified initial influence, which drives a degree-normalized diffusion process using restart-equipped random walks to propagate relevance to contextual neighborhoods. This process prunes irrelevant tokens while preserving user-indicated targets and semantically linked context, distilling a focused yet comprehensive subgraph. Finally, CoreGaze fuses this subgraph with prompt tokens in the frozen large language model decoder, facilitating fine-grained referring generation. Experimental results show that CoreGaze achieves outstanding performance in multiple referring dialogue tasks, showcasing its effectiveness.

Abstract Phishing email attacks have evolved into a significant threat, causing substantial economic and political harm. However, existing detection methods often neglect the data heterogeneity resulting from diverse email sources and are trained on balanced email datasets, which do not accurately reflect real-world scenarios. Meanwhile, with increasing privacy protection regulations, it is crucial to develop methods that enhance phishing email detection capabilities while preserving user privacy. To address these challenges, we propose PhFL, a framework based on heterogeneous federated learning, for detecting phishing emails. PhFL decouples clients’ models into representation learning models and classifiers. The representation learning models can be tailored to clients’ specific needs, and the classifiers are globally shared and re-trained on the server, leveraging the class feature means generated by the representation learning models. Our framework allows each client to leverage its private data locally without providing emails to other clients or the server. The collaboration of class feature means and re-training of classifiers effectively address the challenges of class imbalance and data heterogeneity, enabling improved model performance. Experimental results demonstrate that PhFL outperforms other federated learning methods, particularly when different clients have email datasets from diverse sources and face imbalanced class distributions.

Federated learning, as a distributed machine learning paradigm, enhances privacy protection but faces the challenge of heterogeneity. Data-free knowledge distillation (DFKD) methods attempt to overcome this challenge by using a generator to synthesize samples for fine-tuning a global model. However, these methods often suffer from significant shifts in output distribution, leading to catastrophic forgetting. To tackle these issues, we propose FedFLD, a novel federated forget-less distillation framework that mitigates catastrophic forgetting in DFKD while addressing the heterogeneity challenge. Specifically, FedFLD guides the generator’s training from three key aspects and constrains the output distribution with an elastic weight consolidation penalty term. By synthesizing diverse samples from different perspectives through additional generator updates, FedFLD facilitates effective knowledge distillation from local models to the global model. Additionally, the global model is further optimized via the heterogeneity fine-tuning process, mitigating the bias from heterogeneity and resulting in a more expressive and robust model.

Information Extraction (IE), aiming to extract structured information from unstructured natural language texts, can significantly benefit from pre-trained language models.However, existing pre-training methods solely focus on exploiting the textual knowledge, relying extensively on annotated large-scale datasets, which is labor-intensive and thus limits the scalability and versatility of the resulting models.To address these issues, we propose SKIE, a novel pre-training framework tailored for IE that integrates structural semantic knowledge via contrastive learning, effectively alleviating the annotation burden.Specifically, SKIE utilizes Abstract Meaning Representation (AMR) as a lowcost supervision source to boost model performance without human intervention.By enhancing the topology of AMR graphs, SKIE derives high-quality cohesive subgraphs as additional training samples, providing diverse multi-level structural semantic knowledge.Furthermore, SKIE refines the graph encoder to better capture cohesive information and edge relation information, thereby improving the pre-training efficacy.Extensive experimental results demonstrate that SKIE outperforms state-of-the-art baselines across multiple IE tasks and showcases exceptional performance in few-shot and zero-shot settings.