Reinforcement learning offers a new methodological perspective for the study of the evolution of cooperation. The key lies in state representation, which determines how agents encode environmental information and make decisions accordingly. Existing research has largely focused on whether a specific state representation can promote cooperative behavior, while systematic comparisons of which types of state representations are more conducive to the emergence and maintenance of cooperation remain limited. To address this, the paper proposes a unified memory-one state representation framework to systematically compare different state representations and identify those that more effectively sustain high levels of cooperation. The results show that both the amount of local environmental information contained in the state representation and the way it is organized significantly affect the levels of cooperation. Among the representation methods examined, the one defined by the number of cooperators among an agent and its neighbors, which captures complete local information, proves to be the most effective. However, not all representation methods that include complete local information yield the same effect. Representation methods with overly coarse or overly fine granularity can impair the effective identification of local cooperative structures, leading to the spread of defection. These findings remain robust in different reinforcement learning algorithms, network structures, and game models. This study deepens our understanding of state representation mechanisms and provides guidance for state design in cooperative evolution research using reinforcement learning.

Large Language Model (LLM)-driven multimodal agents are increasingly deployed to execute autonomous tasks via continuous audio interaction. While this paradigm enhances interaction naturalness, it introduces a critical yet under-explored attack surface, as audio inputs inevitably contain environmental noise beyond user control. In this paper, we investigate concurrent audio prompt injection attacks targeting multimodal agents. Distinct from traditional acoustic attacks on voice devices, we propose novel techniques for instruction augmentation and scenario concealment. These methods allow malicious audio instructions to imperceptibly "piggyback" onto user speech, thereby hijacking agents to execute malicious actions. To systematically quantify this threat, we construct AudioAgentSecurity, the first comprehensive benchmark for audio instruction injection attacks, encompassing 8 real-world task scenarios and 10 distinct attack patterns. We evaluate 11 state-of-the-art agents, including Gemini 3 Pro and GPT-4o-audio. Notably, our methods achieve an average Attack Success Rate (ASR) of 69.10% against the advanced Gemini 3 Pro. To counter this threat, we further introduce Cascaded Audio Decoupling and Verification (CADV), a defense mechanism based on source separation and consistency analysis. Compared with existing prompt-level defenses, CADV leverages acoustic source separation and cross-modal consistency analysis to detect audio instruction injections more robustly, achieving over 90% detection success across diverse attack vectors. Finally, real-world experiments with human volunteers on Doubao AI Smartphone in diverse dynamic real-world scenarios confirm the attacks' high stealth and efficacy, while demonstrating that our defense reliably mitigates these vulnerabilities.

As autonomous agents (e.g., OpenClaw) increasingly operate with deep system-level privileges to execute complex tasks, they introduce severe, unmitigated security risks. Existing LLM safety testing methods are largely built around prompt-level inputs and response-level judgments, while recent agent benchmarks remain limited in automation, trajectory coverage, and action-grounded judgment. In this work, we present ASEval, a novel automated security testing framework for autonomous agents. ASEval generates realistic benign multi-turn conversations, perturbs them into risk test cases through carefully designed perturbation operators, executes each case against the target agent, and applies an action-grounded oracle to judge the complete trace. We instantiate ASEval to construct A3S-Bench, a test suite of 2,254 executable multi-turn test cases spanning diverse usage scenarios, delivery patterns, and agent-level security risks. Extensive evaluation on 11 LLM-backed OpenClaw agents confirms that ASEval effectively exposes pervasive security failures. In particular, our proposed perturbation operators raise the average risk-trigger rate from 22.9

Perceptual heterogeneity reflects individuals' different perspectives on the same thing, and competition among individuals with different perceptions of either social dilemmas or payoffs can independently promote cooperation. In this article, we construct a coupled competition mechanism to explore the impact of competition between the two types of perceptions. Specifically, individuals' differing perceptions of social dilemmas are modeled by the snowdrift game (SDG) and the prisoner's dilemma game (PDG), differing perceptions of payoffs are captured by conformists and imitators. Extensive simulations show that the coupled competition mechanism markedly promotes cooperation. PDG players are substantially less competitive than SDG players, which dominate the population. Conformists act as an amplifier, inclined to choose cooperation when cooperators are dominant, thus further promoting cooperation levels. When defectors are dominant, conformists may temporarily choose to defect in the short term, but defection is unsustainable due to lower average accumulated payoffs. We further find that intensified competition among imitators weakens the competitiveness of conformists and diminishes their positive impact on cooperation. Finally, we demonstrate the impact of different proportions of conformists, network structures, and noise levels on the main results to verify the robustness of the model.

Autonomous AI agents tackling Long Horizon Tasks depend on marketplace skills that are certified one at a time: a scanner returns a safety verdict for each skill and declares the ecosystem safe if every package passes. We show that this assumption fails under skill composition. A skill may pass the per-skill scanner individually yet participate in a risky composition when an agent connects its outputs, capabilities, or side effects with those of other scanner-passing skills. This makes skill composition risk a path level property rather than a node level property, explaining why existing skill scanners that inspect individual packages achieve limited interception. To study this threat, we present CompoSkill, a framework that constructs skill composition attacks through a dual attacker system. The white-box attacker knows the victim's installed skill pool and directly injects explicit skill-id sequences; the black-box attacker knows only a role profile, downloads the top marketplace skills for that scenario, builds a Skill Composition Graph, and searches for high risk chains whose implicit lures never name skill identifiers. We further construct CompoSkill-Bench, a benchmark of 1,140 records built from long-horizon professional workflows across five threats and six scenarios on OpenClaw and Nanobot. CompoSkill achieves risk Chain Formation Rates (CFR) up to 83.3

Deep reinforcement learning (DRL) is an effective approach to solving missile guidance problems. However, the broad exploration space and complex environment in the missile guidance process can make it difficult to improve its behavior in the early stage and remain stable in hard scenarios. For this reason, we propose a method named curriculum and imitation-based deep reinforcement learning (CIRL) for missile guidance. CIRL can provide action for the missile controller to hit a randomly maneuvering target with a great intersection angle and overcome detection delay and noise. To bootstrap and stabilize the agent’s training process, CIRL introduces double-adjust imitation learning (DAIL) to help the agent handle both good and bad exploration trajectories. Rule imitation learning is the first part of DAIL, enabling the agent to adapt to traditional guidance laws, avoid continued deterioration, and establish a basic policy in the early iterations. Gaussian self-imitative learning (GSIL) is introduced as the second part, which will help agents attach more importance to well-performed actions. We also apply curriculum learning to reduce the negative effect of imitation learning further promoting the agent’s exploration and enhancing robustness against various factors, including the target’s mobility, delay, and noise. Simulation results validate that CIRL outperforms traditional methods and state-of-the-art DRL-based guidance algorithms with a higher true-hit rate, which reaches 22.7

Neural Architecture Search (NAS) has been an important topic to automate the designs of deep neural networks (DNNs). In this realm, differential NAS via multi-level optimization and zero-shot NAS are perhaps two most popular approaches. However, the existing methods typically suffer from balancing between search efficiency and stability. To address the dilemma, we revisit formulating the differentiable NAS as a single-level optimization problem and propose DeNoiseNAS to address the known coupling issue lying in single-level NAS paradigm. DeNoiseNAS employs a sophisticated search schema that establishes evaluation metrics through theoretical analysis of the neuron optimization process and progressively prunes based on these metrics to identify and remove redundant operators and noisy data samples. As a result, we achieve significant gains regarding search efficiency due to the pruning search space and dataset as well as sub-networks of higher performance due to the elimination of negative impacts from the noisy instances. Numerical experiments across extensive NAS benchmarks well validate the efficacy of DeNoiseNAS. In the DARTS and NAS-Bench-201, while maintaining a competitive search efficiency akin to zero-shot NAS, the architectures uncovered by our strategy surpass the existing state of the art in terms of accuracy, particularly on the ImageNet2012 dataset. In the AutoFormer benchmark, our method efficiently searches for high-performance architectures while consuming fewer resources.

Social norms (collective solutions to social dilemmas) in the real world regulate degrees of wrongdoing, yet classical metanorm games treat violations as binary. We propose the Multivariate Illegal Strategy Metanorm (MISM) model, extending Axelrod's framework to an n-level offense scale with matching gradient punishments, and introduce a reinforcement-learning updater, Multivariate Boldness-Vengefulness Learning (MBVL), that blends Q-learning and evolutionary imitation. Large-scale Monte Carlo simulations on lattice, Watts-Strogatz, and Barab & aacute;si-Albert networks reveal that (i) finer-grained violation levels accelerate the eradication of severe offenders; (ii) larger neighborhood amplify cascading peer punishment; and (iii) graded sanctions achieve higher collective welfare than uniform ones. MISM thus bridges binary-norm models and graded real-world governance, providing a scalable test-bed for studying multilevel norm emergence and compliance. The source code for the MISM model and MBVL algorithm is publicly available at: https://github.com/DAISec-Lab/MISM.

In reality, individuals tend to make different decisions based on differences in relationships and behaviors with their neighbors. Based on this observation, the paper explores the evolution of cooperative behavior when agents develop separated actions for each neighbor by the reinforcement learning approach. Through simulation experiments, it is shown that our model improves the cooperative level compared to results that only consider the agent's own behavior. This is because agents tend to adopt cooperative strategies toward their neighbors while avoiding exploitation, thus promoting the steady expansion of cooperation. Notably, we find that agents do not always choose the action with the highest expected rewards. Therefore, we classify the behavior strategies of the agents into 16 types, corresponding to all possible combinations of actions selected in different states. We observe that agents adopting a specific behavior strategy tend to dominate the evolutionary process: when they choose to cooperate, they switch to defection in the next round regardless of the opponent's action; conversely, when they defect, they switch to cooperation in the next round, again independent of the opponent's behavior. These agents are typically distributed among others with different strategy types, playing a bridging and buffering role. By facilitating the expansion of neighboring agents, they contribute to the spread of cooperative behavior and ultimately enhance the overall level of cooperation in the population. Similar phenomena are also observed under initial specific distributions (e.g., ALLC, ALLD). Next, the hyperparameters of reinforcement learning are analyzed, and the results show that cooperation is easier to maintain and expand when agents make decisions based on past experiences and fully consider potential future rewards. We also compare this model with a control model that adopted the assumption of interactive homogeneity, and further examine the impact of different network structures on the cooperative evolution. Finally, we introduce the memory mechanism of agents as an extended analysis of the model.

This paper investigates the optimal allocation of resources between protection and recovery measures to maximize network robustness when both are available to the defender. Our study finds that as resources increase, a wise defender will quickly shift their focus from recovery to protection, then gradually back to recovery. We have elaborately detailed this nonlinear phenomenon. In addition, our results show that as the cost of defending (including protecting and recovering) critical nodes rises, or the load capacity of network redundancy improves, the defender will allocate a growing share of resources to recovery. The priority assigned to different node classes, i.e., the order of node protection or recovery, also shapes overall resource allocation.

Deep neural networks (DNNs) are suffering from ethical issues such as individual discrimination. In response, extensive NN repair techniques have been developed to adjust models and mitigate such undesired behaviors. However, existing fairness repair methods are typically data-centric, which often lack provable guarantees and generalization to unseen samples. To overcome these limitations, we propose ProF, a novel fairness repair framework with provable guarantees. The key intuition of ProF is to leverage interval bound propagation (a widely used NN verification technique) to soundly capture model outputs over the whole set $\mathcal{S}(\bm{x})$ around a biased sample $\bm{x}$. The derived bounds are utilized to guide fairness repair which encourages the model to produce consistent outputs on $\mathcal{S}(\bm{x})$. Specifically, we integrate fairness constraints and model modifications into a unified constraint-solving formulation, which can be transformed to a Mixed-Integer Linear Programming (MILP) problem solvable by off-the-shelf solvers. The solution to the MILP problem effectively induces a repaired model with guaranteed fairness over the whole set $\mathcal{S}(\bm{x})$. We evaluate ProF on four widely used benchmark datasets and demonstrate that it achieves provable fairness repair, with generalization of up to 95.93% on full datasets and 93.16% on the entire input space. Notably, ProF can be easily configured to support multiple sensitive attributes and more practical fairness definitions, while providing provable repair guarantees and delivering around 90% fairness improvement. Our code is available in this repository.

Graph anomaly detection is significantly challenged by information loss during graph embedding, label scarcity, and intricate structural patterns. While graph neural networks (GNNs) have advanced the field, their efficacy is often limited by sensitivity to data quality and oversmoothing. This paper introduces DARE-G, a novel diffusion-based anomaly detection framework that leverages conditional and unconditional diffusion models to mitigate these limitations. Our approach features a dual-phase diffusion process: conditional anomaly augmentation to synthesize realistic anomalous patterns, and unconditional graph denoising to alleviate information loss and enhance the distinction between normal and anomalous node representations. Extensive experiments on both synthetic and real-world datasets demonstrate DARE-G's significant superiority over state-of-the-art methods, achieving an 8.76 percentage point increase in AUC-ROC compared to the best-performing baselines. Ablation studies further validate the framework's robustness across various GNN backbones, highlighting the critical role of denoising steps in capturing multi-hop structural anomalies. The proposed method establishes new benchmarks for graph anomaly detection, particularly in scenarios with extreme label sparsity and adversarial camouflage.

It is known that deep neural networks may exhibit dangerous behaviors under various security threats (e.g., backdoor attacks, adversarial attacks and safety property violation) and there exists an ongoing arms race between attackers and defenders. In this work, we propose a complementary perspective to utilize recent progress on "neural network repair" to mitigate these security threats and repair various kinds of neural network defects (arising from different security threats) within a unified framework, offering a potential silver bullet solution to real-world scenarios. To substantially push the boundary of existing repair techniques (suffering from limitations such as lack of guarantees, limited scalability, considerable overhead, etc) in addressing more practical contexts, we propose ProRepair, a novel provable neural network repair framework driven by formal preimage synthesis and property refinement. The key intuitions are: (i) synthesizing a precise proxy box to characterize the feature space preimage, which can derive a bounded distance term sufficient to guide the subsequent repair step towards the correct outputs, and (ii) performing property refinement to enable surgical corrections and scale to more complex tasks. We evaluate ProRepair across four security threats repair tasks on six benchmarks and the results demonstrate it outperforms existing methods in effectiveness, efficiency and scalability. For point-wise repair, ProRepair corrects models while preserving performance and achieving significantly improved generalization, with a speedup of 5x to 2000x over existing provable approaches. In region-wise repair, ProRepair successfully repairs all 36 safety property violation instances (compared to 8 by the best existing method), and can handle 18x higher dimensional spaces.

Cooperation is the foundation of social progress, but due to rational individuals often prioritize personal interests, reciprocal cooperation is undermined. The Public Goods Game (PGG) is a classic model for studying group interactions. Traditional PGG assumes a static environment, but in reality, the environment is dynamically changing, and there is an interaction between individual behavior and the environment. Therefore, the stochastic game framework is proposed and applied to study the feedback mechanisms between behavior and the environment. This paper takes the two-state environmental transition mechanism as an example to explore the impact of environmental information perception ability on individual decision-making in the stochastic PGG. Specifically, we use the Q-learning algorithm to depict individual decision-making behavior and consider two types of individuals with different perception abilities: individuals with environmental perception ability select the best action based on the current environmental state, while individuals without environmental perception ability make decisions based solely on historical experience. The experimental results show that environmental information perception significantly lowers the cooperation threshold in the stochastic PGG. By analyzing the microscopic interaction modes of individuals, we find that there is an isolation zone effect between different strategy populations, which effectively prevents the erosion of defection behaviors and ensures the internal stability of cooperation. The extended experiments further validate the robustness of the results. This study shows that environmental information is beneficial for promoting the evolution of cooperation. These findings provide new insights into the cooperation mechanisms in stochastic PGG and offer valuable guidance for promoting cooperation in real-world societies.

Axelrod's model and its subsequent studies have become a valuable framework for fostering cooperation norms among self-interested agents. Within this framework, the concepts of "boldness" and "vengefulness" are specifically employed to characterize agents' behaviors in terms of cooperation and punishment (including metapunishment). Describing behavior solely through the parameters B and V may be overly simplistic and lacks generalizability, making it difficult to apply to other scenarios. Moreover, privacy concerns and the difficulty of evaluating complex states in real-world scenarios limit agents' access to detailed payoff information from their neighbors. To address these questions, our paper employs self-regarding Q-learning, a well-established method for examining the dynamics of strategy updates and agents' learning processes, to investigate whether metanorms can naturally emerge through players' strategy selection. Through extensive experiments, we observe cooperative norms' successful emergence driven by agents' strategy selection variations. Over 90% of agents choose to cooperate on average. In subsequent analyses, we explore the underlying reasons for the emergence of cooperative norms from perspectives of changes in Q-values, punishment and metapunishment frequencies. Additionally, we examine the impact of topological structures on players' strategy selection and assess the emergence of norms across different temptation levels, population sizes, and regulatory intensity levels to validate the model's sensitivity.

Large-scale matrix multiplication is a critical operation in various fields such as machine learning, scientific computing, and graphics processing, but performing it on a single machine introduces significant computational latency. Therefore, matrices are partitioned along different dimensions, decomposed into multiple subtasks, and executed in distributed systems. However, the presence of stragglers in distributed systems can severely impact the speed of matrix multiplication. So, coding schemes are introduced to mitigate the straggler problem. Recently, coding schemes for three-dimensional matrix partitioning have gained increasing attention, including DEP codes. However, these schemes have not focused on decoding accuracy and job completion time. In this paper, with the aim of enhancing decoding accuracy and reducing job completion time, we combine the grouping strategy with Systematic Matdot codes to propose Grouped Systematic Matdot (GSM) codes. Experimental results demonstrate that, compared to DEP codes, GSM codes ensure 100% decoding accuracy and achieve shorter encoding time, communication time, and local computation time, thereby reducing job completion time by at least 45%. Moreover, GSM codes consume fewer memory resources, and as the matrix size increases, their time advantage becomes more pronounced.

Job-seeking is always an inescapable challenge for graduates. It may take a lot of time to find satisfying jobs due to the information gap between students who need satisfying offers and enterprises which ask for proper candidates. Although campus recruiting and job advertisements on the Internet could provide partial information, it is still not enough to help students and enterprises know each other and effectively match a graduate with a job. To narrow the information gap, we propose to recommend jobs for graduates based on historical employment data. Specifically, we construct a heterogeneous information network to characterize the relations between students, enterprises and industries. And then, we propose a meta-path based graph neural network, namely GraphRecruit, to further learn both latent student and enterprise portrait representations. The designed meta-paths connect students with their preferred enterprises and industries from different aspects. Also, we apply genetic algorithm optimization for meta-path selection according to application scenarios to enhance recommendation suitability and accuracy. To show the effectiveness of GraphRecruit, we collect five-year employment data and conduct extensive experiments comparing GraphRecruit with 4 classical baselines. The results demonstrate the superior performance of the proposed method.

Social manufacturing has significantly advanced the industrial Internet of Things (IIoT), integrating information technology and operation technology to enhance production efficiency and quality, and to foster new business models. This integration, however, introduces novel risks, including advanced persistent threats, which demand robust security measures to safeguard IIoT systems. This article proposes an iCyberGuard game model, tailored for IIoT environments, designed to imitate the cyber and physical attacks for information and operation technologies. Then, we used a reinforcement learning algorithm to compute the optimal strategy. We conducted comprehensive simulation experiments, which demonstrate that our model the strategic interactions between attackers and defenders. Participants are enabled to learn adaptively, discerning optimal strategies based on the intelligence of their adversaries. Finally, we explain the practical significance of the best strategy of defenders or attackers, and how users can rely on these best strategies to strengthen the security performance of the network.

Most individuals opt for vaccination to acquire immunity protection and prevent disease transmission. However, individuals cannot obtain perfect immunity protection after vaccination, due to various factors such as the limitation of vaccine itself, storage and transportation. Failed vaccination experiences can alter individuals' perception of vaccination behavior. To analyze the influence of vaccine efficacy on vaccination behavior with adaptive perception, we propose a novel vaccination game model. The results demonstrate that for the moderate vaccination cost, the introduction of adaptive perception can promote vaccination behavior, and the promoting effect becomes more pronounced in the population with smaller perception fluctuation. Nonetheless, vaccination behavior is still constrained by a significant number of free-riders when vaccine effectiveness is high. Analyzing the distribution of strategies among individuals with different degrees, it is revealed that the reduction in vaccinated individuals influenced by free-riders predominantly occurs in individuals with low-degree. Furthermore, we examine the coupled effects of vaccination cost and vaccine efficacy on vaccination behavior, considering various levels of perception fluctuations. The results indicate the crucial role of vaccination cost in enhancing vaccination behavior, and previous findings also are consistent across scenarios with diverse vaccination cost. Our work contributes to an improved comprehension of vaccination behavior considering vaccine efficacy and perception.

Model inversion attacks (MIAs) aim to recover private data from inaccessible training sets of deep learning models, posing a privacy threat. MIAs primarily focus on the white-box scenario where attackers have full access to the model's structure and parameters. However, practical applications are usually in black-box scenarios or label-only scenarios, i.e., the attackers can only obtain the output confidence vectors or labels by accessing the model. Therefore, the attack models in existing MIAs are difficult to effectively train with the knowledge of the target model, resulting in sub-optimal attacks. To the best of our knowledge, we pioneer the research of a powerful and practical attack model in the label-only scenario. In this paper, we develop a novel MIA method, leveraging a conditional diffusion model (CDM) to recover representative samples under the target label from the training set. Two techniques are introduced: selecting an auxiliary dataset relevant to the target model task and using predicted labels as conditions to guide training CDM; and inputting target label, pre-defined guidance strength, and random noise into the trained attack model to generate and correct multiple results for final selection. This method is evaluated using Learned Perceptual Image Patch Similarity as a new metric and as a judgment basis for deciding the values of hyper-parameters. Experimental results show that this method can generate similar and accurate samples to the target label, outperforming generators of previous approaches.