Protein-Protein Interactions (PPIs) prediction is crucial for understanding cellular functions and disease mechanisms. Existing deep learning–based methods primarily rely on direct interaction within the PPI network to update protein representations. However, (1) such networks overlook the potential associations between functionally similar proteins, limiting the smoothing capability of Graph Neural Networks (GNNs) in learning representations for similar nodes. (2) Additionally, most approaches fail to adequately model the latent dependencies among interaction types (edge labels), which hinders their performance in PPI prediction tasks. To address these limitations, we propose TELC-PPI, a topology-enhanced and label correlation-aware model for protein-protein interactions prediction. Specifically, TELC-PPI first identifies similar proteins by leveraging both the topological information of the PPI network and the label distributions of nodes, constructing similarity edges. Then, it incorporates label co-occurrence statistics into the learning of label embeddings. Experimental results on multiple datasets and under various data split settings demonstrate that TELC-PPI significantly outperforms existing methods, validating the effectiveness of our model design.

LLM-based multi-agent systems exhibit strong collaborative capabilities but often suffer from redundant communication and excessive token overhead. Existing methods typically enhance efficiency through pretrained GNNs or greedy algorithms, but often isolate pre- and post-task optimization, lacking a unified strategy. To this end, we present SafeSieve, a progressive and adaptive multi-agent pruning algorithm that dynamically refines the inter-agent communication through a novel dual-mechanism. SafeSieve integrates initial LLM-based semantic evaluation with accumulated performance feedback, enabling a smooth transition from heuristic initialization to experience-driven refinement. Unlike existing greedy Top-k pruning methods, SafeSieve employs 0-extension clustering to preserve structurally coherent agent groups while eliminating ineffective links. Experiments across benchmarks (SVAMP, HumanEval, etc.) showcase that SafeSieve achieves 94.01% average accuracy while reducing token usage by 12.4%-27.8%. Results further demonstrate robustness under prompt injection attacks (1.23% average accuracy drop). In heterogeneous settings, SafeSieve reduces deployment costs by 13.3% while maintaining performance. These results establish SafeSieve as an efficient, GPU-free, and scalable framework for practical multi-agent systems. Our code can be found below.

Shuang Liu, Ruijia Zhang, Ruoyun Ma, Yujia Deng, Lanyi Zhu, Jiayu Li, Zelong Li, Zhibin Shen, Mengnan Du. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

In this paper, we study the problem of (finite sum) minimax optimization in the Differential Privacy (DP) model. Unlike most of the previous studies on the (strongly) convex-concave settings or loss functions satisfying the Polyak-Lojasiewicz condition, here we mainly focus on the nonconvex-strongly-concave one, which encapsulates many models in deep learning such as deep AUC maximization. Specifically, we first analyze a DP version of Stochastic Gradient Descent Ascent (SGDA) and show the utility bound in terms of the Euclidean norm of the gradient for the empirical risk function. We then propose a new method with less gradient noise variance and improve the upper bound to the best-known result for DP Empirical Risk Minimization with non-convex loss. We also discussed several lower bounds of private minimax optimization. Finally, experiments on AUC maximization, generative adversarial networks, and temporal difference learning with real-world data support our theoretical analysis.

The goal of the Inverse reinforcement learning (IRL) task is to identify the underlying reward function and the corresponding optimal policy from a set of expert demonstrations. While most IRL algorithms’ theoretical guarantees rely on a linear reward structure, we aim to extend the theoretical understanding of IRL to scenarios where the reward function is parameterized by neural networks. Meanwhile, conventional IRL algorithms usually adopt a nested structure, leading to computational inefficiency, especially in high-dimensional settings. To address this problem, we propose the first two-timescale single-loop IRL algorithm under neural network parameterized reward and provide a non-asymptotic convergence analysis under overparameterization. Although prior optimality results for linear rewards do not apply, we show that our algorithm can identify the globally optimal reward and policy under certain neural network structures. This is the first IRL algorithm with a non-asymptotic convergence guarantee that provably achieves global optimality in neural network settings.

Image-text matching is an important and challenging task in the field of multimedia analysis, aimed at bridging the semantic gap between visual content and language descriptions. Although this field has significant implications for enhancing cross-modal interaction, most previous work still faces challenges in accurately aligning images with text descriptions, especially when dealing with images containing rich semantic information. To address this issue, we propose a novel context-aware image-text matching model that extracts and summarizes visual region information aligned with multiple text descriptions from a single image. Specifically, we designed an adaptive context-aware self-attention module to extract representations of visual regions and text words. By controlling the complementary semantic relationships within each modality, our model can adaptively capture contextual information for each modality. We then introduce a Transformer-based encoder layer to extract features at multiple levels and aggregate region-level features into image-level features. Finally, fine-grained word-region alignment is conducted to match image features with corresponding text features. To evaluate the effectiveness of our method, we conducted extensive experiments on two benchmark datasets, Flickr30K and MS-COCO. The experimental results show that our model outper-forms several state-of-the-art baselines in image-text matching tasks, demonstrating the effectiveness and practicality of our model in handline cross-modal retrieval tasks.

Negative sampling (NS) is widely used in knowledge graph completion (KGC) to generate negative triples for contrastive learning during training. However, existing NS methods are not suitable when multimodal information is incorporated into KGC models. Due to their complex design, these methods are also inefficient. In this paper, we propose the integration of Modality-Aware Adversarial Training (IMAT) to generate higher-quality negative samples for Multimodal Knowledge Graph Completion (MMKGC), and introduce a Relation-Enhanced Cross-modal Attention (RECA) mechanism to evaluate bidirectional attention weights between multimodal features using relational information, thereby improving the model's ability to identify hard negative samples. Our approach represents a joint design of MMKGC models and training strategies, surpassing 16 recent MMKGC methods and achieving new state-of-the-art results on three public MMKGC benchmarks.

In recent years, cross-modal image-text retrieval has gained significant attention due to its ability to efficiently retrieve semantically relevant information from large-scale multimedia data. The rise of deep learning has provided new perspectives for addressing the heterogeneity challenge in cross-modal retrieval. However, existing hashing methods often struggle to balance efficient retrieval with fine-grained semantic alignment and global semantic understanding, thereby limiting retrieval accuracy. To tackle these challenges, this paper proposes a novel deep multi-feature hashing network (DMFHN), designed to achieve both efficient and fine-grained cross-modal retrieval through compact binary hash codes. The core of DMFHN lies in the synergy between the feature optimization encoder and a bidirectional GRU. The feature optimization encoder integrates self-attention mechanisms with depthwise separable convolutions to effectively capture both global dependencies and local details in images. Specifically, the module first employs multi-head self-attention to model the global contextual information of an image, then utilizes depthwise separable convolutions to extract crucial local features. By fusing global relationships with local details, it generates a more expressive image representation. Meanwhile, the bidirectional GRU enhances textual features by capturing sequential dependencies and contextual semantics within the text. Additionally, we design a cross-modal feature fusion strategy that dynamically integrates image and text representations, further improving fine-grained semantic expressiveness. Through a well-optimized hashing function, DMFHN constrains and quantizes multi-modal features into compact hash codes, ensuring high retrieval efficiency while maintaining strong cross-modal semantic consistency. Finally, experiments conducted on MIRFLICKR-25K and NUS-WIDE, two real-world datasets, demonstrate that DMFHN achieves state-of-the-art performance in image-text retrieval, significantly outperforming existing mainstream methods.

Knowledge graph completion (KGC) aims to learn high-quality representations of entities and relations to predict missing entities in knowledge graphs (KGs). Multimodal knowledge graph completion (MMKGC) enhances entity representation by incorporating multimodal information such as images. However, existing methods struggle to effectively filter out irrelevant features and fail to fully exploit relational dependencies, thereby limiting their reasoning capability. To address these issues, we propose a Relationship-driven Collaborative and Integrative Multimodal Knowledge Graph Completion model (RCIM-KGC), which leverages relationship-driven mechanisms to enhance multimodal knowledge representation. Specifically, the Relationship-driven Collaborative Module (RDCM) adaptively selects and integrates visual and structural information based on relational importance to optimize entity representation. Meanwhile, the Relationship-driven Integrative Module (RDIM) computes relational weights for neighboring entities to guide the aggregation of multimodal neighborhood information, thereby improving reasoning performance. Experimental results demonstrate that RCIM-KGC outperforms existing multimodal KGC methods across multiple benchmark datasets, validating its effectiveness.