Molecular optimization in drug discovery aims to discover molecules with improved target properties, but practical lead optimization often requires more than high predicted scores. A useful candidate should also be actionable: it should be reachable from known molecules through valid local structural transformations, so that it can be interpreted as a plausible revision within an evolving chemical series. Existing de novo and single-molecule optimization methods do not explicitly model such reachability, especially when both the target molecules and the intermediate molecules connecting them to known compounds are unknown. In this work, we formulate actionable molecular optimization as sequential expansion of a molecule-transfer graph, where nodes are molecules and edges encode valid local transformations. We propose MolWorld, a molecule world model-guided framework that treats the current molecule-transfer graph as an evolving search state. At each iteration, MolWorld selects local anchor contexts, generates candidate molecules conditioned on these contexts, evaluates their properties, and uses a learned world model to update the evolving molecule world by retaining admissible candidates and inserting them into the molecule-transfer graph. The expanded molecule world then guides subsequent optimization. Experiments on property optimization and docking-based tasks show that MolWorld discovers high-property molecules while maintaining substantially stronger structural connectivity, supporting actionable and sequential molecular design.

Large-scale networks of agents are increasingly applied to software engineering, scientific analysis, web automation, organizational workflows, and social simulation, yet existing multi-agent architectures lack a unified framework to explain why some designs scale to long-horizon, multi-step tasks while others fail. Recent open agent ecosystems such as OpenClaw and agent-only social environments such as Moltbook further highlight that large, partially open-ended agent interaction is no longer merely a laboratory setting, but an emerging deployment regime with new coordination, safety, and governance challenges. As these systems grow, their behavior is fundamentally shaped by how agents are connected, how information is stored, and how states are updated over time. In this survey, we introduce a hierarchical taxonomy of agent systems along three core dimensions-architecture topology (centralized vs. decentralized), memory scope (global vs. local), and update behavior (static vs. dynamic)-which together induce eight system categories that organize prior work and make architectural trade-offs explicit. Using this taxonomy, we analyze how design choices influence scalability, coordination efficiency, communication overhead, planning depth, and robustness under partial failure, and we identify common failure modes and open challenges, including consistency management, agent routing, federation boundaries, and stability under noise or disruption.

Vision-Language Models often struggle with complex visual reasoning due to the visual information loss in textual CoT. Existing methods either add the cost of tool calls or rely on localized patch-based embeddings that are insufficient to extract semantics in multi-step reasoning. We propose "Decompose, Look, and Reason" (DLR), a reinforced latent reasoning framework that dynamically decomposes queries into textual premises, extracts premise-conditioned continuous visual latents, and deduces answers through grounded rationales. We introduce a three-stage training pipeline and propose a novel Spherical Gaussian Latent Policy to enable effective exploration in the latent space. Extensive experiments on vision-centric benchmarks show that DLR consistently outperforms strong baselines, including text-only, interleaved multimodal CoT, and latent reasoning methods, while providing superior stepwise interpretability.

Graph Neural Networks (GNNs) have achieved notable success in graph-structured data tasks, such as social networks, recommendation systems, molecular structures, and citation networks. By leveraging a message-passing framework, GNNs capture node dependencies and demonstrate strong performance across diverse tasks. However, most existing GNN models neglect edge directionality in directed graphs, limiting their ability to accurately model information propagation and node dependencies. To this end, this paper proposes the Directed Aggregation Graph Convolutional Network (DAGCN), a novel approach that incorporates directional features from both first-order and second-order neighbors, effectively capturing directional information in directed graphs. DAGCN extends the receptive field with a single-layer aggregation mechanism, mitigating challenges caused by symmetry in traditional methods. Additionally, a parameterized weighting scheme dynamically adjusts the fusion of local and global information, enhancing the node representation and graph structure modeling accuracy. The flexible neighbor aggregation strategy further improves adaptability and robustness. Extensive experiments on several publicly available datasets show that DAGCN outperforms existing methods in capturing directional information and expanding the receptive field. On the CITESEER-FULL dataset, DAGCN achieves 95.63 +/- 0.44% accuracy, surpassing GCN by 6.39 percentage points (89.24 +/- 0.75%). On the SQUIRREL dataset, DAGCN improves by 22.02 percentage points, reaching 75.87 +/- 2.16% compared to GCN's 53.85 +/- 2.21%. These results underscore the importance of incorporating directional information to enhance performance, especially in directed graph tasks.

Semantic segmentation of high-resolution remote sensing imagery plays a critical role in applications such as disaster assessment. However, deploying large models on Autonomous Aerial Vehicles (AAVs) remains challenging due to inherent conflicts among accuracy, model size, and computational efficiency. To address these challenges, we propose FMC-ULite, a novel lightweight architecture designed to achieve a better balance between accuracy and efficiency for real-time processing. Our model incorporates four key innovations, including a Fast Fourier Transform (FFT)-based fusion module for enhanced edge feature extraction and noise suppression in the frequency domain, a simplified MobileNetV3-Large encoder that substantially reduces parameter count, a cross-layer feature fusion (CLFF) module to effectively integrate multi-scale semantic and detail information, and an attention-gated decoder with multi-scale dilated convolutions to prioritize critical disaster regions. Furthermore, an adaptive combined loss function is introduced to alleviate class imbalance. Experiments conducted on the RescueNet dataset show that our model achieves competitive accuracy compared to advanced lightweight methods under a comparable parameter budget, demonstrating its strong suitability for real-time disaster assessment using AAVs.

With the rapid development of lunar exploration, devices deployed on the lunar surface will encounter emergency events including meteor impacts and lunar dust storms. Additionally, the extreme lunar environment characterized by intense radiation and the high latency of Earth-based cloud computing pose severe challenges to real-time and reliable task processing. Existing collaborative computing and fault-tolerant schemes are not fully efficient in dynamic radiation environments. To address these challenges, a collaborative computing architecture integrating lunar surface devices and lunar orbit satellites is proposed for lunar emergency tasks. Comprehensive network, radiation, communication and computing models are established to support this architecture. The problem is then formulated as a multi-objective optimization problem and solved by the Radiation-aware hiErarchical collAborative Fault-Tolerant Reinforcement Learning (REAFTRL) algorithm, which integrates radiation-aware, hierarchical decision-making, and fault-tolerant execution with feedback mechanisms. Simulations show the proposed collaborative computing scheme outperforms traditional fault-tolerant strategies in task completion time, completion rate, and error rate, providing a reliable solution for future lunar exploration.

Spatio-temporal time series forecasting plays a crucial role in many real-world applications, particularly in traf fic flow prediction. Numerous methods have been successfully deployed in traffic management systems and have achieved promising performance. However, most existing approaches focus primarily on modeling spatiotemporal dependencies, while overlooking intrinsic characteristics of traffic data, such as temporal periodicity and semantic similarities among different spatial regions. To address this limitation, we propose a novel Spatio-Temporal Aware Hybrid Sequence Model (STAHSM) based on an encoder-decoder framework. The proposed model learns spatio-temporal embedding vectors to represent inherent traffic properties and captures spatiotemporal information from three perspectives: enhancing latent representations via vector concatenation in the encoder, introducing a spatio-temporal feature-enhanced Graph Convolutional Recurrent Unit (GCRU) unit in the decoder, and learning a periodic graph to model dynamic interactions among spatial nodes. Extensive exper iments conducted on four benchmark datasets demonstrate that STAHSM consistently achieves state-of-the-art performance.

AC Optimal Power Flow (ACOPF) and Security-Constrained Unit Commitment (SCUC) are fundamental optimization problems in power system operations. ACOPF serves as the physical backbone of grid simulation and real-time operation, enforcing nonlinear power flow feasibility and network limits, while SCUC represents a core market-level decision process that schedules generation under operational and security constraints. Although these problems share the same underlying transmission network and physical laws, they differ in decision variables and temporal coupling, and prior learning-based approaches address them in isolation, resulting in disjoint models and representations.We propose a learning framework that jointly models ACOPF and SCUC through a shared graph-based backbone that captures grid topology and physical interactions, coupled with task-specific decoders for static and temporal decision-making. Training includes solver supervision with physics-informed objectives to enforce AC feasibility and inter-temporal operational constraints. To evaluate generalization, we assess cross-case transfer on unseen grid topologies for ACOPF and SCUC without retraining, and systematic generalization on the UC-ACOPF problem using unsupervised, physics-based objectives and a power-dispatch consensus mechanism. Experiments across multiple grid scales demonstrate improved performance and transferability relative to existing learning-based baselines, indicating that the model can support learning across heterogeneous power system optimization problems.

Large language models (LLMs) have become widely adopted as automated judges for evaluating AI-generated content. Despite their success, aligning LLM-based evaluations with human judgments remains challenging. While supervised fine-tuning on human-labeled data can improve alignment, it is costly and inflexible, requiring new training for each task or dataset. Recent progress in auto prompt optimization (APO) offers a more efficient alternative by automatically improving the instructions that guide LLM judges. However, existing APO methods primarily target text-only evaluations and remain underexplored in multimodal settings. In this work, we study auto prompt optimization for multimodal LLM-as-a-judge, particularly for evaluating AI-generated images. We identify a key bottleneck: multimodal models can only process a limited number of visual examples due to context window constraints, which hinders effective trial-and-error prompt refinement. To overcome this, we propose BLPO, a bi-level prompt optimization framework that converts images into textual representations while preserving evaluation-relevant visual cues. Our bi-level optimization approach jointly refines the judge prompt and the I2T prompt to maintain fidelity under limited context budgets. Experiments on four datasets and three LLM judges demonstrate the effectiveness of our method.

Understanding how information or influence propagates through a network, such as during an epidemic outbreak or the spread of misinformation, is a fundamental yet challenging problem. While prior works have focused on cascade prediction (forecasting future infected nodes), network inference (recovering latent global diffusion graphs), or source localization (identifying diffusion's origin), these approaches do not recover the actual "who-infected-whom" propagation tree for a specific diffusion instance. We introduce DIPT (Deep Identification of Propagation Trees), a probabilistic framework that infers propagation trees from final observed node diffusion states, without knowledge of the underlying diffusion mechanism. DIPT models local influence strengths between nodes and uses a discrete-continuous alternating optimization strategy to jointly learn the diffusion mechanism and infer the propagation structure. Empirical results across eight real-world datasets demonstrate that DIPT consistently outperforms existing approaches in reconstructing propagation trees.

High pass rates on established programming benchmarks such as HumanEval and LiveCodeBench do not always show whether a model can reason about algorithms. Many fixed benchmarks eventually become part of the public training ecosystem through released problem statements, editorials, and generated solutions, allowing later models to improve partly by exposure rather than by stronger algorithmic ability. We introduce ALGOBENCH, a framework that automatically builds novel algorithmic problems from known competitive-programming problems through structured constraint-shifting transformations. Each accepted ALGOBENCH variant is traceable to a source problem, but must make the original reference algorithm fail. Beyond pass@k, we introduce complexity-aware metrics – including OPTT, OPTS, TRAPRATE, GAPT, and CONSENS – to test whether a solution is not only functionally correct but also asymptotically suitable for the generated problem. Experiments across multiple LLMs and prompting strategies show that performance drops sharply on ALGOBENCH variants, retrieval can increase reuse of the old algorithm, and many correct-looking solutions fail to meet the required complexity. Error analysis shows that failures are mainly algorithmic rather than implementation-level, suggesting that ALGOBENCH evaluates adaptation beyond functional correctness.

Cognitive diagnosis (CD) and knowledge tracing (KT) are two core tasks for assessing learner states in intelligent educational systems. Although joint modeling studies have attempted to incorporate diagnostic information into the KT process, most approaches employ loosely coupled designs or rely on implicit utilization of cognitive cues, making it challenging to establish a principled, structured connection between static cognitive assessment and dynamic sequence prediction. This article proposes a dual-phase cognitive processes modeling framework (DPF) that explicitly connects CD and KT through a cognitive prior constraint mechanism. In Phase I, the model fuses multidimensional interactive data, including emotional features, to infer learners' cognitive states, yielding continuous and interpretable cognitive representations. In Phase II, these representations initialize the hidden states of the sequence model, imposing cognitive constraints on the sequential modeling of learners' response behaviors to guide response correctness prediction on specific knowledge components. Experiments conducted on multiple real-world educational datasets show that the DPF achieves competitive performance compared with classical CD models and representative KT methods in both diagnostic capability and prediction accuracy. Ablation studies further indicate that the observed improvements are mainly related to the explicit incorporation of cognitive prior constraints rather than solely increased model complexity. Overall, this study proposes a cognition-guided modeling paradigm that effectively bridges static cognitive assessment and dynamic learning prediction within a unified framework.

As low Earth orbit (LEO) satellite constellations expand, the volume of on-orbit data processing increases significantly. However, these systems face challenges in data processing due to heterogeneous data distribution and limited onboard resources. This article presents an innovative framework for personalized federated learning (PFL) tailored to heterogeneous LEO satellite networks, which mitigates data processing challenges and optimizes distributed computational performance across the satellite constellation. Our approach introduces personalized models built on individual satellite datasets, coupled with dynamic model aggregation and pruning techniques for efficient training. By overcoming data heterogeneity and localizing global models, our method significantly improves performance over traditional approaches. Extensive simulation validates the effectiveness of our PFL framework, which demonstrates its superiority in integrating federated learning (FL) with model pruning in satellite networks. Simulation results show that our proposed PFL method outperforms four comparative algorithms, which highlights its effectiveness in addressing the unique challenges of LEO satellite networks.

Traditional computation offloading and resource allocation strategies encounter several issues that lead to poor service experience and resource wastage. The resource allocation scheme lacks the flexibility to adapt to the time-varying offloading demands of User Equipment (UEs). Furthermore, there is an imbalance between UEs seeking better service and Service Providers (SPs) aiming to minimize cost expenditures. In this paper, we propose a knowledge-defined networking-based Multi-Layer Computation Offloading and Resource Allocation strategy optimization (ML-CORA) architecture. Based on the ML-CORA, we design a Multi-Layer Local-Global Collaborative computation offloading and resource allocation strategy optimization (ML2GC) algorithm. The basic level of the ML2GC algorithm expresses and optimizes computation offloading demands from the perspective of UE (local), while the meta level optimizes the resource allocation strategy on demand from the perspective of the SP (global), achieving a collaborative multi-objective optimization for a win-win system between UEs and SPs. The two-layer structure of the ML2GC algorithm outputs continuous and discrete actions respectively, which improves the flexibility and efficiency of the algorithm while effectively balancing the interests of all parties and promoting efficient resource utilization. Simulation results based on the real-world dataset of Shanghai Telecom indicate that the ML2GC algorithm significantly improves both social welfare and resource utilization compared to baseline algorithms.

Network tomography is a crucial problem in network monitoring, where the observable path performance metric values are used to infer the unobserved ones, making it essential for tasks such as route selection, fault diagnosis, and traffic control. However, most existing methods either assume complete knowledge of network topology and metric formulas-an unrealistic expectation in many real-world scenarios with limited observability-or rely entirely on black-box end-to-end models. To tackle this, in this paper, we argue that a good network tomography requires synergizing the knowledge from both data and appropriate inductive bias from (partial) prior knowledge. To see this, we propose Deep Network Tomography (DeepNT), a novel framework that leverages a path-centric graph neural network to predict path performance metrics without relying on predefined hand-crafted metrics, assumptions, or the real network topology. The path-centric graph neural network learns the path embedding by inferring and aggregating the embeddings of the sequence of nodes that compose this path. Training path-centric graph neural networks requires learning the neural netowrk parameters and network topology under discrete constraints induced by the observed path performance metrics, which motivates us to design a learning objective that imposes connectivity and sparsity constraints on topology and path performance triangle inequality on path performance. Extensive experiments on real-world and synthetic datasets demonstrate the superiority of DeepNT in predicting performance metrics and inferring graph topology compared to state-of-the-art methods.

Data collection and distribution through crowdsensing has become an emerging trend in smart city scenarios. By leveraging existing vehicle resources without deploying dedicated infrastructure, Vehicular CrowdSensing (VCS) provides low-cost and high-mobility data collection on road networks. Typically, the Crowdsensing Platform (CP) issues data collection tasks, recruits Sensing Vehicles (SVs) to complete tasks, and sells the collected data to Data Demanders (DDs). Here, the goal of CP is to maximize profits through data collection and sales, and the goal of DDs is to improve satisfaction by purchasing high-quality sensing data. However, due to individual rationality and selfishness, not all SVs are willing to complete the sensing task. Therefore, how to motivate SVs to complete sensing tasks with high quality and efficiency, while handling the relationship among CP, DDs, and SVs, is a problem that needs to be considered. To solve the above problems, this paper proposes a Reputation-based Hybrid Incentive Approach (RHIA), with the goal of maximizing the utility of CP, SVs, and DDs. Specifically, in order to improve the task completion quality of SVs, we introduce vehicle reputation to measure SVs. Then, we propose a one-toone bargaining game between CP and each SV, and use the reputation value as the sequential basis of the game. we also design a unique SV Trajectory Planning Algorithm (STPA). Finally, based on real-world datasets, the effectiveness of our proposed RHIA and STPA is verified. Our proposed method can ensure the long-term stability of the VCS system, which also improves the utility of participating individuals.

Mobile Edge Computing (MEC)-assisted Space-Air-Ground Integrated Network (SAGIN) has emerged as a promising solution to achieve wide-area coverage, ubiquitous connectivity, and pervasive edge computing. However, limited on-board resources, rapid motion of satellites and UAVs, and highly dynamic ground workloads challenge traditional static and single-layer task scheduling, making it difficult to balance delay and energy consumption. To address these challenges, this paper proposes a task prediction-driven UAV trajectory planning, hierarchical pre-caching, and adaptive offloading (TP${}^{3}$Of) scheme. Distinct from traditional decoupled approaches, we establish a prediction-driven deep coupling framework. Specifically, we first design a dual-scale prediction mechanism using Markov chains and attention-based ConvLSTM to forecast task distribution. Based on this, we employ a training-free Monte Carlo Tree Search (MCTS) algorithm to execute proactive UAV trajectory planning for improving coverage efficiency. Subsequently, a coverage and delay-aware hierarchical pre-caching strategy is developed to optimize content placement across space-air-ground layers. Finally, we propose a weight-adaptive Deep Q-Network (DQN) that dynamically adjusts reward weights based on real-time energy and delay constraints to optimize offloading decisions. Extensive experimental results clearly highlight that the TP${}^{3}$Of scheme significantly outperforms the others, achieving superior performance in both delay and energy consumption.

Large language models (LLMs) are revolutionizing various fields due to their powerful generation capabilities. However, their immense computational complexity poses significant challenges in resource consumption, inference latency, and data privacy for traditional cloud-centric deployments. Edge artificial intelligence (Edge-AI) offers promising LLMs deployment solutions by leveraging distributed resources at the network edge. However, existing approaches struggle to adapt to dynamic workloads and efficiently utilize heterogeneous resources in Mobile Edge Computing (MEC) environments. This paper proposes a Dynamic Batching and Adaptive Partitioning (DyBAP) scheme for LLMs deployment, which utilizes ubiquitous geo-distributed resources via end-edge-cloud collaboration. Firstly, we formulate a collaboration deployment optimization problem to minimize inference latency and resource usage under heterogeneous resource and user requirements for latency and accuracy constraints, which is NP-hard. Secondly, to solve this, we develop a dynamic batch fusion optimization algorithm that optimizes the batch size of inference by utilizing the parallel processing power of computing units to balance the latency and resource usage. A block-aware partition optimization algorithm based on multi-agent reinforcement learning (MARL) is proposed for efficient transformer block allocation, integrating mobility awareness for optimal partitioning across dynamic network environments. Simulation results demonstrate the superiority of DyBAP over other benchmarks, reducing inference latency by 17.94% and saving 11.12% in memory resource consumption compared to the end-edge-cloud collaboration approaches.

Tool-augmented language models can extend small language models with external executable skills, but scaling the tool library creates a coupled challenge: the library must evolve with the planner as new reusable subroutines emerge, while retrieval from the growing library must remain within a fixed context budget. Existing tool-use and skill-library methods typically treat tools as flat or text-indexed memories, causing prompt cost to grow with library size and obscuring the typed, compositional structure of executable code. We propose CoCoDA, a framework that co-evolves the planner and tool library through a single code-native structure: a compositional code DAG. Nodes are primitive or composite tools, edges encode invocation dependencies, and each node stores a typed signature, description, pre/post-condition specification, and worked examples. At inference time, Typed DAG Retrieval prunes candidates by symbolic signature unification, ranks survivors by descriptions, filters them by behavioral specifications, and disambiguates with examples, keeping expensive context materialization on progressively smaller candidate sets. At training time, successful trajectories are folded into validated composite tools, while the planner is updated with a DAG-induced reward that credits composites by their primitive expansion size. We provide theoretical results showing retrieval cost reduction, sublinear retrieval time, compositional advantage under the shaped reward, monotone co-evolution under conservative updates, and DAG well-formedness. Across mathematical reasoning, tabular analysis, and code task benchmarks, CoCoDA enables an 8B student to match or exceed a 32B teacher on GSM8K and MATH and consistently improves over strong tool-use and library-learning baselines.

AC optimal power flow (ACOPF) is foundational yet computationally expensive in power grid operations, driving learning-based surrogates for large-scale grid analysis. These surrogates, however, often fail to generalize across network topologies, a critical gap for deployment on grids not seen during training and for routine operational what-if studies. We introduce LUMINA-Bench, a comprehensive benchmark suite for ACOPF surrogate learning covering multi-topology pretraining, transfer, and adaptation. The benchmark evaluates homogeneous and heterogeneous architectures under single- and multi-topology learning settings using unified metrics that capture both predictive accuracy and physics-informed constraint violations. We additionally compare constraint-aware training objectives, including MSE, augmented Lagrangian, and violation-based Lagrangian losses, to characterize accuracy-robustness trade-offs across settings. Data processing, training, and evaluation frameworks are open-sourced as the LUMINA suite to support reproducibility and accelerate future research on feasibility-aware OPF surrogates.