Generative Retrieval (GR) offers a promising paradigm for recommendation through next-token prediction (NTP). However, scaling it to large-scale industrial systems introduces three challenges: (i) within a single request, the identical model inputs may produce inconsistent outputs due to the pagination request mechanism; (ii) the prohibitive cost of encoding long user behavior sequences with multi-token item representations based on semantic IDs, and (iii) aligning the generative policy with nuanced user preference signals. We present GenRec, a preference-oriented generative framework deployed on the JD App https://www.jd.com that addresses above challenges within a single decoder-only architecture. For training objective, we propose Page-wise NTP task, which supervises over an entire interaction page rather than each interacted item individually, providing denser gradient signal and resolving the one-to-many ambiguity of point-wise training. On the prefilling side, an asymmetric linear Token Merger compresses multi-token Semantic IDs in the prompt while preserving full-resolution decoding, reducing input length by ~2× with negligible accuracy loss. To further align outputs with user satisfaction, we introduce GRPO-SR, a reinforcement learning method that pairs Group Relative Policy Optimization with NLL regularization for training stability, and employs Hybrid Rewards combining a dense reward model with a relevance gate to mitigate reward hacking. In month-long online A/B tests serving production traffic, GenRec achieves 9.5% improvement in click count and 8.7% in transaction count over the existing pipeline.

Early fault detection of degraded gears at different speeds is both essential and challenging. Adaptive blind deconvolution methods have shown considerable promise for extracting fault characteristics from complex vibration signals. Their performance depends on accurate cyclic frequency estimation and optimal filter length selection. However, this estimation often fails due to gear meshing shock interference and early weak fault characteristics. Additionally, determining the filter length relies on additional metrics with inefficient search strategies, thereby limiting the overall reliability and efficiency. To address these issues, an adaptive blind deconvolution via convolutional neural network (ABDCNN) is proposed. First, we employ an envelope harmonic product spectrum guided by gear frequency-domain features to reduce interference from noise and meshing shocks, enabling precise estimation of the target cyclic frequency. Then, an attention mechanism is integrated into the convolutional neural network to jointly optimize filter coefficients and length estimation, thereby improving computational efficiency. Simulations and gear contact fatigue experiments demonstrate that ABDCNN enables more efficient detection of early faults across different speeds while maintaining strong interpretability.

Reconfigurable intelligent surface (RIS) has emerged as a promising solution to enable ultra-massive multiple-input multiple-output (MIMO) for 6G wireless communications. To reduce pilot overhead, multi-resolution codebook-based beam training has been proposed for RIS-aided communication systems. However, the hardware constraints of the RIS such as the limited control capability, mutual coupling, and manufacturing tolerances have not been fully considered, which may result in non-ideal beam patterns and misleading beam search directions, thereby degrading the overall beam training performance. To address this issue, in this paper, we revisit the multi-resolution codebook design and beam training scheme by taking the RIS hardware constraints into account. First, we propose a wide-beam generation method via multi-beam superposition, where the RIS reflection coefficients are optimized to achieve high beam gain with minimal fluctuations. Then, a posterior-based beam training scheme is proposed to adaptively correct erroneous beam search directions, thereby improving the beam training accuracy. We implement a prototype RIS with 32x32 reflective elements and deploy a 26 GHz millimeter-wave RIS-aided wireless communication testbed to experimentally evaluate the proposed scheme. The results indicate that our approach achieves low training overhead and a data rate close to the exhaustive search method.

As a key feature that has the potential to enable many advanced applications, the integration of sensing functionality is considered essential in the 6G network, which motivates the design of integrated sensing and communication (ISAC) systems. However, traditional ISAC systems based on non-overlapped resource allocation face the challenge of poor energy and spectral efficiency. In this paper, we implement an ISAC system named Horus based on reconfigurable intelligent surfaces, which provides an energy-efficient solution to sense objects in a wide range of blind areas. With a carefully designed ISAC protocol, Horus can transmit sensing information to the receiver with high spectral efficiency. We have verified the ability of the proposed system in two case studies of multi-modal sensing and around-corner radar early warning, respectively. Our demonstration video can be found in [11].

MOTIVATION: Many types of networks, such as co-expression or ChIP-seq-based gene-regulatory networks, provide useful information for biomedical studies. However, they are often too full of connections and difficult to interpret, forming "indecipherable hairballs." RESULTS: To address this issue, we propose that a Bayesian network can summarize the core relationships between gene expression activities. This network, which we call the LatentDAG, is substantially simpler than conventional co-expression network and ChIP-seq networks (by two orders of magnitude). It provides clearer clusters, without extraneous cross-cluster connections, and clear separators between modules. Moreover, one can find a number of clear examples showing how it bridges the connection between steps in the transcriptional regulatory network and other networks (e.g. RNA-binding protein). In conjunction with a graph neural network, the LatentDAG works better than other biological networks in a variety of tasks, including prediction of gene conservation and clustering genes. AVAILABILITY AND IMPLEMENTATION: Code is available at https://github.com/gersteinlab/LatentDAG.

The reconstruction of subsurface ocean temperature using sea surface observations and in situ Argo measurements is an important yet challenging task. The availability of long-term and high-resolution sea surface remote sensing, combined with advancements in deep learning technology, has opened new opportunities for studying subsurface temperature (ST) reconstruction. In this study, a novel spatial–temporal Siamese convolutional neural network (SSCNN) is proposed to improve the accuracy of ST reconstruction in the Indian Ocean. First, considering the distinctions of temperature characteristics among different sea areas, a multiscale division scheme based on the correlation coefficient of integral ST is designed for refined reconstruction modeling. Second, since ocean heat is significantly affected by solar radiation, asymmetric convolutional operation with rectangular patches and kernels is designed to capture the information characteristics in longitude and latitude directions, respectively. Third, given the temporal changes and correlations of ocean temperature, an SSCNN with shared parameters is proposed for multiview feature mining and accurate temperature structure reconstruction. The reconstructed results provide a precise depiction of the subsurface Indian Ocean dipole (sub-IOD)’s evolution, including the spatial distribution of positive and negative anomaly signals and its temporal changes. It demonstrates that the subsurface dipole index series obtained from SSCNN reconstruction is consistent with that from International Pacific Research Center (IPRC) observation, remaining within a reasonable error range. Comparative experiments indicate that the SSCNN model surpasses other existing methods in terms of higher accuracy and smaller error. Overall, this study provides a promising approach for effectively reconstructing the ST using deep learning methods and offers valuable insights for analyzing the evolution of subsurface positive dipole in Indian Ocean.

Driven by the increasingly serious air pollution problem, nowadays different systems can be used to achieve the monitoring task of air quality index (AQI) in urban areas. In this article, we design a novel unmanned aerial vehicle-aided (UAV-aided) AQI monitoring system, called AQ360, which detects the air quality level from the 360-degree aerial panoramic images taken by the onboard camera. Specifically, we first present our own AQI recognition approach based on the physical form of the haze pictures, where the AQI is jointly decided by the images captured along six directions over the target location. Then, we study the UAV placement problem of selecting UAV's flight altitude and 2-D coordinates during the monitoring process. The objective is to save the system energy consumption while maintaining the accuracy of estimating AQI distribution. For practical considerations, we implement and evaluate the proposed system in real-world scenarios. The results show that our system can provide a lower AQI recognition error compared with existing vision-based monitoring approaches, and energy consumption is also reduced when applying for large-area tasks.

While integrated electronic systems (IESs) are widely used in military and civilian applications, their security issues are barely studied. By analyzing the architecture of the system and the characteristics of bus communication, this paper proposes an intrusion detection method based on the message sequence and behavioral rules of subsystems. According to the bus protocol, messages are divided into periodic and aperiodic messages. For the previous, we adopt sequence analysis and propose an algorithm that extract the sequence intelligently to determine if there are anomalies. For aperiodic messages, we detect the anomalies by modeling the system behaviors as decision trees. Through implementing experiments on our simulation system, we demonstrate that the proposed detection is more accurate than the existing schemes while incurring both lower false negative rate and lower false positive rate.