Graphical user interface (GUI) agents are widely used in general-purpose digital automation. Large language models, vision-language models, and multimodal foundation models can now interpret screenshots, follow natural-language instructions, and execute grounded actions in websites, mobile applications, and desktop operating systems. However, despite their different execution conditions, these settings are commonly grouped under one label. This survey reviews representative frameworks, models, datasets, and benchmarks of GUI agents through a common technical stack: observation, grounding, planning, memory, execution, and verification. We advocate the treatment of web, mobile, and desktop/operating-system agents as distinct operational regimes with different observabilities, action semantics, hidden-state dependences, execution costs, and side-effect risks. We also describe a shift from next-action prediction to dependable execution with increased focus on outcome verification, execution efficiency, and trustworthiness. The survey concludes with future directions in belief-state tracking, adaptive sensing, hybrid GUI-tool execution, and human oversight for reliable computer use.

Large language models (LLMs) have achieved strong performance on many language tasks, but they still struggle with culturally grounded symbolic reasoning. Existing benchmarks have not systematically evaluated the progressive capability chain required for cross-cultural understanding, which involves intra-cultural symbolic understanding, cross-cultural symbolic alignment, and cross-cultural conflict identification. To address this gap, we propose CCU-Bench, a systematic benchmark for evaluating cross-cultural understanding in LLMs. Grounded in Hofstede’s cultural dimensions theory, CCU-Bench focuses on three culturally distinct contexts, China, Japan, and Mexico, and is organized around three dimensions, image, connotation, and emotion. The benchmark is constructed through three stages, data collection, question generation, and quality control, resulting in a high-quality evaluation set of 3029 question-answer pairs in five formats. Experiments on 19 mainstream LLMs show that current models perform unsatisfactorily on this task, achieving an average score of 57.8%. Closed-source models consistently outperform open-source models, while cross-cultural symbolic alignment remains the most challenging sub-task. Further error analysis reveals that the dominant failures stem from deficiencies in cultural knowledge, biases in intent mapping, and weak higher-order reasoning about cross-cultural conflicts, rather than simple instruction-following issues. These findings highlight persistent limitations of current LLMs in culturally grounded reasoning and demonstrate that CCU-Bench provides a standardized benchmark for culturally aware artificial intelligence research.

Chinese medical named entity recognition (CMNER) is a fundamental task in medical information extraction. It is crucial for building downstream applications, such as clinical knowledge graphs, and enabling intelligent clinical decision-making. However, existing mainstream approaches, including lexicon-enhanced, span-based, and grid-based tagging methods, struggle with the absence of natural boundaries, complex nested entity structures, and long-range contextual dependencies inherent in clinical texts. To address these challenges, we propose a novel dual-stream fusion with dynamic grid interaction (D2GI) model, which performs deep semantic mining by integrating complementary feature streams and adaptive grid interactions to accurately capture complex entity boundaries and inter-character relations. Specifically, our dual-stream fusion architecture leverages RoFormer to extract long-range dependencies and incorporates Word2Vec to provide stable prior semantics, thereby enhancing the representation of rare medical terms. Furthermore, to overcome the limitations of static refinement, the dynamic grid interaction module employs a gated attention mechanism to adaptively fuse local and global contexts, facilitating accurate recognition of nested entities. Multiple experiments on three public datasets demonstrate that D2GI is superior to state-of-the-art baselines, achieving F1-score improvements of 1.48 percentage points (PPs) on CMeEE-V2, 1.61 PPs on DiaKG, and 3.08 PPs on CCKS2020.

In response to the increasing demand for heterogeneous data interaction and cross-disciplinary modeling in aerospace situational awareness simulations, we propose a microservice platform for data-model computing (MP-DMC), which is a high-performance microservice framework built on a container cloud. The proposed MP-DMC framework unifies the control of data and models through a distributed node resource management and scheduling strategy, integrating an election-optimized leader-follower mechanism, a predictive model based on a double-moving-average long short-term memory (DMA-LSTM) network for dynamic elastic scaling, and an intelligent load migration algorithm to address management inefficiencies, prevent node crashes, and mitigate resource oscillations under high-concurrency conditions. Experimental results demonstrate that the proposed MP-DMC framework outperforms mainstream algorithms in terms of election performance, node scaling efficiency, task response time, and load balancing, including consensus algorithms (Paxos, Raft, and PBFT), elastic scaling methods (HPA, DMA-HPA, ProSmart HPA, and RL), and scheduling algorithms (round-robin, purely random, and weighted random), achieving exceptional resource allocation performance and system availability.

This paper addresses the problem of end-effector position-tracking control for micro-nano free-floating space robots in Cartesian space without relying on explicit analytical kinematic or dynamic models. To address this challenge, we develop a two-layer learning architecture. In the first layer, a deep neural network is used for kinematic learning to capture the nonlinear mapping from end-effector Cartesian coordinates to joint angular velocities and to generate reference joint trajectories. In the second layer, a Koopman-operator-based network is employed to construct an approximately linearized representation of the joint-space dynamics of free-floating space robots. Based on this model, we propose a terminal fractional-order model predictive control scheme that incorporates the Grünwald-Letnikov fractional-order operator, thereby enhancing online control performance and improving tracking speed and accuracy relative to conventional model predictive control. Simulation results verify the effectiveness of the proposed method, demonstrating accurate and rapid end-effector trajectory tracking, all without requiring explicit analytical kinematic and dynamic models in the controller design, while the training pipeline relies solely on input-output trajectories.

This paper introduces a novel knee exoskeleton designed to provide assistive torque to the knee joint and facilitate postoperative rehabilitation. Drawing inspiration from the anatomy of the human knee joint, the exoskeleton incorporates a biomimetic structure that integrates the “optimal axis of rotation” with the trajectory of the instantaneous center of rotation. This design effectively mitigates the issue of kinematic misalignment in the human-machine joint, offering a more comfortable wearing experience for users. Furthermore, a highly integrated quasi-direct drive system is incorporated into the knee exoskeleton. The drive system can deliver a continuous torque of 18 N·m and a peak torque of 45 N·m to sufficiently meet the daily walking needs of patients. From a control perspective, a unified sliding mode control strategy based on iterative learning control is proposed to maintain high tracking accuracy. This strategy enables the system to achieve accurate tracking of the knee motion trajectory by continuously updating the control inputs through successive gait cycles. Experimental results demonstrate that the designed biomimetic joint mechanism effectively resolves the issue of misalignment in the human-machine joint. Furthermore, the synergistic control strategy demonstrates excellent steady-state precision and transient robustness in terms of trajectory tracking during human-exoskeleton cooperative walking.

Conventional circuit systems typically exhibit limited capability in suppressing disturbances with strong temporal structures. This study proposes a bio-inspired anti-interference framework based on neural synchronization mechanisms. Hodgkin-Huxley (HH) and Izhikevich (IZH) networks with different topologies are compared in terms of synchronization stability, energy deviation, and hardware feasibility under external interference. The proposed framework integrates the neuron models and topological structures of neural networks to establish a unified evaluation system for bio-inspired anti-interference circuits. The simulation results demonstrate that the HH network exhibits stronger robustness under strong coupling, maintaining a synchronization index near 1.0 and a lower relative energy deviation than the IZH network. The small-world topology achieves the best disturbance tolerance, whereas the chain topology exhibits the weakest robustness for both models. In addition, hardware implementation demonstrates that the IZH network significantly reduces lookup table consumption, and both the HH and IZH circuits maintain rhythmic firing and waveform integrity under 20 mV square-wave interference. These results indicate that neural synchronization mechanisms provide an effective and practically implementable mechanism for adaptive anti-interference circuit design.