Multimodal large language models (MLLMs) recently showed strong capacity in integrating data among multiple modalities, empowered by generalizable attention architecture. Advanced methods predominantly focus on language-centric tuning while less exploring multimodal tokens mixed through attention, posing challenges in high-level tasks that require fine-grained cognition and emotion understanding. In this work, we identify the attention deficit disorder problem in multimodal learning, caused by inconsistent cross-modal attention and layer-by-layer decayed attention activation. To address this, we propose a novel attention mechanism, termed MOdular Duplex Attention (MODA), simultaneously conducting the inner-modal refinement and inter-modal interaction. MODA employs a correct-after-align strategy to effectively decouple modality alignment from cross-layer token mixing. In the alignment phase, tokens are mapped to duplex modality spaces based on the basis vectors, enabling the interaction between visual and language modality. Further, the correctness of attention scores is ensured through adaptive masked attention, which enhances the model's flexibility by allowing customizable masking patterns for different modalities. Extensive experiments on 21 benchmark datasets verify the effectiveness of MODA in perception, cognition, and emotion tasks.

With social bots evolving continually, the new bots post highly anthropomorphic posts to evade detection. For post content information, existing methods choose to extract features from single posts or use the rough characterization of the pretraining language model for bot detection. However, the evolution of bots has made them better at camouflage in previous posts processing techniques. According to our observation, the purpose of bot posting is to publicize different contents in different periods, and its posting often shows abnormal interest changes, while human posting is to express hobbies and daily life, and its interest changes are relatively stable. Therefore, we extract the interest changes between multiple posts published by users for bot detection. In this paper, we propose a social Bot detection model based on the Window Strategy(BotWS). Specifically, we first employ the window strategy to obtain the user posting windows, each containing multiple posts. Then, we extract the interest changes between the posting windows by multi-head attention mechanism. Finally, we embed the interest changes information into the user representation and construct a heterogeneous graph classification module to classify. We conduct experiments on three challenging datasets. Results indicate that our approach achieves state-of-the-art.

Abstract This work presents a space‐frequency prefiltering scheme for slowly time‐varying TDD MC‐CDMA downlink communications with multiple antennas at the base station (BS). Unlike the conventional spatially uncorrelated block fading channel model, both channel variation in each packet and spatial correlation are considered in the design. In the TDD mode, the mobile terminals (MTs) transmit training signals at the end of each uplink packet. In the following downlink packet, the BS computes the signal weights on different antennas and subcarriers for each MT in each symbol period based on the channel state predicted from the received training signals. The goal is to minimize the total required transmit power while keeping the received signal‐to‐interference‐plus‐noise ratio (SINR) as the target for each MT. Moreover, the maximum packet length for satisfying the SINR requirements has been determined. The results indicate that the total required transmit power can be reduced by a lower mobile speed or more BS antennas. As a result, the maximum packet length can be extended in virtue of the power reduction. Copyright © 2008 John Wiley & Sons, Ltd.