agent

This paper examines the decision-making processes of physicians and intelligent agents within the healthcare sector, particularly focusing on their characteristics, architectures, and approaches. We provide a theoretical insight into the evolving role of artificial intelligence (AI) in healthcare, emphasizing its potential to address various healthcare challenges. Defining features of intelligent agents are explored, including their perceptual abilities and behavioral properties, alongside their architectural frameworks, ranging from reflex-based to general learning agents, and contrasted with the rational decision-making structure employed by physicians. Through data collection, hypothesis generation, testing, and reflection, physicians exhibit a nuanced approach informed by adaptability and contextual understanding. A comparative analysis between intelligent agents and physicians reveals both similarities and disparities, particularly in adaptability and contextual comprehension. While intelligent agents offer promise in enhancing clinical decisions, challenges with types of dataset biases pose significant hurdles. Informing and educating physicians about AI concepts can build trust and transparency in intelligent programs. Such efforts aim to leverage the strengths of both human and AI toward improving healthcare delivery and outcomes.

No abstract available

Abstract Multi-agent systems enable the division of complicated tasks into individual objects that can cooperate. Such architecture can be useful in building solutions in the Internet of Medical Things (IoMT). In this paper, we propose an architecture of such a system that ensures the security of private data, as well as allows the addition and/or modification of the used classification methods. The main advantages of the proposed system are based on the implementation of blockchain technology elements and threaded federated learning. The individual elements are located on the agents who exchange information. Additionally, we propose building an agent with a consortium mechanism for classification results from many machine learning solutions. This proposal offers a new model of agents that can be implemented as a system for processing medical data in real-time. Our proposition was described and tested to present advantages over other, existing state-of-the-art methods. We show, that this proposition can improve the Internet of Medical Thing solutions by presenting a new idea of a multi-agent system that can separate different tasks like security, or classification and as a result minimize operation time and increase accuracy.

Abstract The consequences of natural or man-made catastrophes can be devastating. To minimize its impact, it is crucial to carry out a rapid analysis of the affected environment in the moments after they occur, especially from the perspective of alert notification or crisis management. In this context, the use of UAVs, understood as the technological basis on which intelligent systems capable of providing support to rescue teams is built, has positively contributed to face this challenge. In this article the design of a multi-agent architecture which enables the deployment of systems made up of intelligent agents that can monitor environments affected by a catastrophe and provide support to human staff in the decision-making process is proposed. These environments, known in advance, are characterized through a set of points of interests that are critical from the point of view of aerial surveillance and monitoring. To conduct an intelligent information analysis, a formal model of normality analysis is employed, which makes possible the definition of surveillance components. These represent the knowledge bases of the agents responsible for monitoring environments. Likewise, the architecture envisages communication and cooperation mechanisms between the different agents, as the basis for fusing information to assess the overall level of risk of the monitored environment. A case study is presented in which the spread of toxic smoke in an industrial complex which has just suffered a hypothetical earthquake is monitored.

No abstract available

No abstract available

No abstract available

No abstract available

No abstract available

No abstract available

No abstract available

No abstract available

No abstract available

No abstract available

No abstract available

No abstract available

No abstract available

The North Atlantic Treaty Organization (NATO) Research Task Group IST-152 developed a concept and a reference architecture for intelligent software agents performing active, largely autonomous cyber-defense actions on military assets. The group released a detailed report, briefly reviewed in this article, where such an agent is referred to as an Autonomous Intelligent Cyber-defense Agent (AICA). In a conflict with a technically sophisticated adversary, NATO military networks will operate in a heavily contested battlefield. Enemy malware will likely infiltrate and attack friendly networks and systems. Today’s reliance on human cyber defenders will be untenable on the future battlefield. Instead, artificially intelligent agents, such as AICAs, will be necessary to defeat the enemy malware in an environment of potentially disrupted communications where human intervention may not be possible. The IST-152 group identified specific capabilities of AICA. For example, AICA will have to be capable of autonomous planning and execution of complex multi-step activities for defeating or degrading sophisticated adversary malware, with the anticipation and minimization of resulting side effects. It will have to be capable of adversarial reasoning to battle against a thinking, adaptive malware. Crucially, AICA will have to keep itself and its actions as undetectable as possible, and will have to use deceptions and camouflage. The report identifies the key functions and components and their interactions for a potential reference architecture of such an agent, as well as a tentative roadmap toward the capabilities of AICA.

Nowadays, network users' demands for low latency services and individual data security grow rapidly. In this article, we conceive a new communication network architecture based on the concept of intelligent agents. The proposed intelligent agent serves as a virtual secretary of the corresponding user to request desired service from Internet companies or neighboring agents, which can reduce Internet surfing latency and prevent privacy leakage. Moreover, to improve the forwarding efficiency of data packages, a new mixed routing structure and the AI router are also introduced, to support various on-demand services with low latency. Overall, the proposed communication network architecture is expected to be more efficient than the existing one, and inspire readers to deliberate the revolution of future communications networks.

Future networks such as 6G are projected to incorporate in-network intelligence toward achieving a zero-touch network. In this regard, several approaches proposed for the organizational architecture of future networks. Similar to microservice-based service design, a multi-agent-based network automation architecture was proposed as a competing paradigm for service-oriented architecture. The proposed architecture outlines the design guideline for intelligent network systems using agents as atomic and autonomous service units that can be used as building blocks. As a continuation of this approach, we design a Quality of Service (QoS) agent to control and manage stringent services such as remote surgery. QoS agent is intelligent that can capture and respond proactively to network traffic and workload distribution, showing hourly and seasonal patterns. QoS agents can be used as a building block along with traffic classification and traffic prediction agents for an intelligent networking system. For evaluation, a campus network is designed using a NetSim environment considering a three-tier network architecture. The QoS agent dynamically finds the best path for a particular service depending on the requirements. The agent communicates with the traffic prediction agent and traffic classifier agent to collect information about the network and services. Moreover, the QoS agent also observes the network states. Using these values along with existing network topology knowledge, it ranks the available paths to proactively allocate the best path for a service. Evaluation results suggest that with the appropriate accuracy of traffic prediction, the proposed approach can autonomously adapt in allocating a path for a given service.

We consider the problem of multi-agent navigation and collision avoidance when observations are limited to the local neighborhood of each agent. We propose InforMARL, a novel architecture for multi-agent reinforcement learning (MARL) which uses local information intelligently to compute paths for all the agents in a decentralized manner. Specifically, InforMARL aggregates information about the local neighborhood of agents for both the actor and the critic using a graph neural network and can be used in conjunction with any standard MARL algorithm. We show that (1) in training, InforMARL has better sample efficiency and performance than baseline approaches, despite using less information, and (2) in testing, it scales well to environments with arbitrary numbers of agents and obstacles. We illustrate these results using four task environments, including one with predetermined goals for each agent, and one in which the agents collectively try to cover all goals. Code available at https://github.com/nsidn98/InforMARL.

No abstract available

No abstract available

No abstract available

No abstract available

No abstract available

Over the last few years, machine learning and data mining methods (MLDM) are constantly evolving, in order to accelerate the process of knowledge discovery from data (KDD). Today's challenge is to select only the most relevant knowledge from those extracted. The present paper is directed to these purposes, by developing a new concept of knowledge mining for meta-knowledge extraction, and extending the most popular machine learning methods to extract meta-models. This new concept of knowledge classification is integrated on the cognitive agent architecture, so as to speed-up its inference process. With this new architecture, the agent will be able to select only the actionable rule class, instead of trying to infer its whole rule base exhaustively.

No abstract available

No abstract available

No abstract available

No abstract available

No abstract available

No abstract available

This paper presents a model to define heterogeneous agents that solve problems by sharing the knowledge retrieved from the AND cooperating among them. The control structure of those agents is based on a general purpose Multi-Agent architecture (SKELETONAGENT) based on a deliberative approach. Any agent in the architecture is built by means of several interrelated modules: control module, language and communication module, skills modules, knowledge base, yellow pages, etc. … The control module uses an agenda to activate and coordinate the agent skills. This agenda handles actions from both the internal goals of the agent and from other agents in the environment. In the paper, we show a high level agent model, which is later instantiated to build a set of heterogeneous specialized agents. The paper describes how SKELETONAGENT has been used to implement different kinds of agents and a specialized Multi-Agent System (MAS). The implemented MAS, MAPWEB-ETOURISM, is the specific implementation of a general WEB gathering architecture, named MAPWEB, which extends SKELETONAGENT. MAPWEB has been designed to solve problems in WEB domains through the integration of information gathering and planning techniques. The MAPWEB-ETOURISM system has been applied to a specific WEB domain (e-tourism) which uses information gathered directly from several WEB sources (plane, train, and hotel companies) to solve travel problems. This paper shows how the proposed architecture allows to integrate the different agents tasks with AI techniques like planning to build a MAS which is able to gather and integrate information retrieved from the WEB to solve problems.

No abstract available

No abstract available

Abstract The paper presents the formalism of an intelligent decision-making system based on multi-agent neurocognitive architectures, which has an architectural similarity to the human brain. An invariant of the organizational and functional structure of the intellectual decision-making process based on the multi-agent neurocognitive architecture is developed. An algorithm for teaching intelligent decision-making systems based on the self-organization of the invariant of multi-agent neurocognitive architectures is presented. Using this algorithm, an intelligent agent was trained and the architecture of the learning process was built on the basis of an invariant of neurocognitive architecture. Further research is related to training an intelligent agent in more complex behavior and expanding the capabilities of an intelligent decision-making system based on multi-agent neurocognitive architectures.

No abstract available

With the rapid development of underwater robots, underwater communication techniques, etc., the Autonomous Underwater Vehicle (AUV) cluster network has emerged as a candidate paradigm to perform underwater civil and military applications, e.g., underwater target tracking. In this paper, we focus on how to utilize networking and multi-agent artificial intelligence technique to improve underwater target tracking. In particular, to improve the flexibility and scalability of the AUV cluster network, we employ Software-Defined Networking (SDN) and Centralized Training with Decentralized Execution (CTDE)-based Multi-Agent Reinforcement Learning (MARL) technologies, to propose a Hierarchical Software-Defined Multiple AUVs Reinforcement Learning (HSD-MARL) framework. For the MARL mechanism in HSD-MARL, we propose an advantage-attention mechanism and present the architecture of Multi-AUV Advantage-Attention Actor-Critic (MA-A3C), to address slow convergence and poor scalability issues on the AUV cluster network of large-scale. Further, to improve the utilization rate of advantage samples especially when the MA-A3C is utilized to perform AUV cluster network-based underwater tracking, we propose an ‘advantage resampling’ method based on experience replay buffer. Evaluation results showcase that our proposed approaches can perform exact underwater target tracking based on AUV cluster network systems and outperform some recent research products in terms of convergence speed, tracking accuracy, etc.

The multiple Autonomous Underwater Vehicle (AUV)-assisted cooperative system or the AUV-based Underwater Ad-hoc Networks (UAN) system has been considered as a highly-potential future in underwater data surveillance. In this paper, we propose grid-based distributed data collection architecture and define two categories of navigation modes. Based on the proposed data collection model, we propose MADAC, a scheme based on AUV-based UAN to cooperatively collect data from 6G-driven underwater wireless networks. We utilize the Software-Defined Networking (SDN) technique to re-organize the architecture of AUV-based UAN and propose software-defined actor-critic MARL framework. Based on the proposed MARL framework, we present the paradigm of MADDPG algorithm with optimal similarity attention mechanism (MADDPG-SA), to plan the paths for the AUV-based UAN, especially the cooperative underwater obstacle avoidance, the task distribution balancing, the Value of Information (VoI) are concurrently taken into account. In particular, the proposed MADDPG-SA improves the running efficiency of the proposed MADDPG-SA by encouraging the agent to learn from the similar and better-performance agent. The evaluation results demonstrate that the proposed MADAC can schedule the AUV-based UAN to perform efficient underwater data collection, reduce data collection time and energy consumption, and balance data collection tasks in the AUV-based UAN.

No abstract available

Increasingly, artificial intelligence (AI) is being used to support automotive systems, including autonomous vehicles (AVs) with self-driving capabilities. The premise is that learning-enabled systems (LESs), those systems that have one or more AI components, use statistical models to make better informed adaptation decisions and mitigate potentially dangerous situations. These AI techniques largely focus on uncertainty factors that can be explicitly identified and defined (e.g., environmental conditions). However, the unexpected behavior of human actors is a source of uncertainty that is challenging to explicitly model and define. In order to train a learning-enabled AV, developers may use a combination of realworld monitored data and simulated external actor behaviors (e.g., human-driven vehicles, pedestrians, etc.), where participants follow defined sets of rules such as traffic laws. However, if uncertain human behaviors are not sufficiently captured during training, then the AV may not be able to safely handle unexpected behavior induced by human-operated vehicles (e.g., unexpected sudden lane changes). This work introduces a non-cooperative game theory and reinforcement learning-based (RL) framework to discover and assess an AV's ability to handle high-level uncertain behavior(s) induced by human-based rewards. The discovered synthetic data can then be used to reconfigure the AV to robustify onboard behaviors.

Software-defined vehicles (SDVs) leverage vehicle-to-everything (V2X) communication to enable advanced connectivity and autonomous driving capabilities. However, this increased interconnectivity also exposes them to cyber threats such as spoofing, denial-of-service attacks, and data manipulation, making intrusion detection systems (IDS) essential for ensuring SDV security and reliability. In this work, we propose a novel intrusion mitigation approach that integrates Advantage Actor-Critic (A2C) reinforcement learning with a Long Short-Term Memory (LSTM) network to detect anomalies and intrusions in V2X communications. The LSTM component captures temporal dependencies in V2X data, enhancing the model's ability to identify emerging attack patterns, while the A2C framework dynamically adjusts defensive actions, including flagging, blocking or monitoring traffic, based on evolving threat levels. Experimental results demonstrate the model's effectiveness, achieving high detection accuracy and sensitivity. Additionally, we analyze how the system adapts over time, becoming more confident in its decision-making and optimizing security enforcement. This work enhances SDV cybersecurity by introducing a learning-based adaptive intrusion response system aiming at mitigating threats in highly dynamic vehicular networks.

The integration of artificial intelligence in computer vision tasks has significantly transformed computing technology, especially in applications such as autonomous vehicles, medical image diagnosis, and video surveillance systems. However, multi-object tracking, a key component of modern computer vision, faces challenges such as object occlusion, changing object appearance, and complex interactions with the background. This paper discusses the improvements in multi-object tracking achieved by integrating multi-agent reinforcement learning with modern tracking methods. A novel tracking system is proposed that uses the deep deterministic policy gradient algorithm for multi-agent systems in combina- tion with the YOLO v7 object detector. The system uses centralized learning with decentralized execution and the Actor- Critic architecture to dynamically and accurately track multiple objects in real time. The proposed method improves the tracking performance by improving the coordination between agents and efficiently handling the dynamics of the environ- ment. Experiments conducted on the MOT16 dataset showed significant improvements in tracking accuracy and processing speed, making this method suitable for a variety of applications.

No abstract available

The exploitation/utilization of marine resources and the rapid development of urbanization along coastal cities result in serious marine pollution, especially underwater diffusion pollution. It is a non-trivial task to detect the source of diffusion pollution, such that the disadvantageous effect of the pollution can be reduced. With the vision of 6G framework, we employ Autonomous Underwater Vehicle (AUV) flock and introduce the concept of AUV-based network. In particular, we utilize the Software-Defined Networking (SDN) technique to update the controllability of the AUV-based network, leading to the paradigm of SDN-enabled multi-AUVs network Intelligent Transportation Systems (SDNA-ITS). For SDNA-ITS, we utilize artificial potential field theories to model the control model. To optimize the system output, we introduce the graph-based Soft Actor-Critic (SAC) algorithm, i.e., a category of Multi-Agent Reinforcement Learning (MARL) mechanism where each AUV can be regarded as a node in a graph. In particular, we improve the optimization model based on Centralized Training Decentralized Execution (CTDE) architecture with the assistance of the SDN controller, by which each AUV can efficiently adjust its speed towards the diffusion source. Further, to achieve exact path planning for detecting the diffusion source, a dynamic detection scheme is proposed to output the united control policy to schedule the SDNA-ITS dynamically. Simulation results demonstrate that our approaches are available to detect the underwater diffusion source when the actual scenario is taken into account and perform better than some recent research products.

With the rapid development of the underwater internet of things (UIoT), underwater security challenges, especially the problem of illegal invasion, are becoming increasingly prominent, threatening the security of underwater environments and infrastructures. In this article, we propose an approach combining software defined networking (SDN) and multi-agent deep reinforcement learning (MADRL) to improve the efficiency of cooperative tracking of Autonomous Underwater Vehicle (AUV) systems in complex underwater environments. By leveraging SDN technology, this novel approach achieves centralized control and management of multi-AUV systems. Then, we construct a trajectory prediction network based on the attention mechanism and long short-term memory (LSTM) to generate the target trajectories and provide prior knowledge for collaborative pursuit. Finally, a novel MADRL method combining bidirectional LSTM and multiagent soft actor-critic is proposed to make real-time, accurate, distributed, and adaptive pursuit decisions when a part of AUVs break down. The experimental results demonstrate that the proposed methods can pursue the target successfully at the fastest speed, as compared to the latest MADRL methods.

Safety assurance of autonomous vehicles (AVs) is particularly challenging when considering the infinite number of scenarios an AV may encounter. As such, existing scenario generation approaches optimize search to derive dangerous refinements of a (same) abstract scenario given as input. In this paper, we propose a scenario generation approach that derives dangerous (collision-inducing) concrete scenarios from arbitrary abstract scenarios (under reasonable assumptions). As added novelty, our approach allows to compare the level of danger offered by different abstract scenar-ios. We evaluate the collision avoidance capacity of the Transfuser AV controller by generating, then simulating, collision-inducing 2-actor scenarios at a road junction. Results show that distinctions at higher abstraction levels yield measurable differences in simulation.

Model-free control has gained in popularity in recent years for its adaptive, data-driven solutions to complex control objectives. This paper discusses the application of model-free reinforcement learning (RL) to the control of multibody aerial vehicles. First, a summary of model-free RL as it applies to robotic systems is presented. Then, a model-free RL controller is demonstrated, whereby an actor-critic algorithm is used to stabilize the swing of a slung payload carried by an autonomous quadcopter aerial vehicle, utilizing the latter’s full continuous action-space. Its effectiveness and adaptability are first demonstrated in simulation, and then validated on an experimental testbed. The algorithm is shown to be computationally efficient enough to adapt to the experimental testbed in real-time, and to work within the framework of widely-used autopilot software ArduPilot.

The rapid development of intelligent underwater devices promotes marine exploitation activities, including marine resource exploitation, marine target tracking, etc. This work will present how to utilize the Autonomous Underwater Vehicle (AUV) swarm or multi-AUVs system to track the underwater diffusion pollution, especially the equipotential line of particular concentration. Different from most of the current research, in this work, we take the AUV swam as a network system and utilize the Software-Defined Networking (SDN) technique to optimize the network architecture, constructing an SDN-enabled AUV network Intelligent Transportation Systems (ITS). With the centralized management ability of the SDN technique, we propose the software-defined Centralized Training Decentralized Execution (CTDE) architecture based on the graph-based Soft Actor-Critic (SAC) algorithm to optimize the system control and management. To improve the computing and training efficiency, we embed the self-attention mechanism into the critic network construction, leading to a self-attention-based SAC algorithm. Evaluation results demonstrate that our proposed approach is able to exactly track the equipotential lines of a particular concentration in many categories (with different types of equipotential lines (including the shape, noise, and diffusion value)) of underwater diffusion fields. Meanwhile, our proposed approaches outperform some classical schemes in system awards, tracking errors, etc.

No abstract available

No abstract available

No abstract available

No abstract available

No abstract available

Traditional approaches for steering a vehicle using machine vision require large amounts of robust hand-crafted software which is both time consuming and expensive. The presented method uses a deep neural network to teach cars to steer themselves without any additional software. We created a labeled dataset for the ACTor (Autonomous Campus TranspORt) electric vehicle by pairing real world images taken during a drive with the associated steering wheel angle. We trained a model end to end using modern deep learning techniques including convolutional neural networks and transfer learning to automatically detect relevant features in the input and provide a predicted output. This means that no traditional hand engineered algorithm features were required for this implementation. We currently use an pretrained inception network on the ImageNet dataset to leverage the high level features learned from ImageNet to the steering problem through transfer learning. We removed the top portion of the network and replaced it with a linear regression node to provide the output. The model is trained end to end using backpropagation. The trained model is integrated with vehicle software on ROS (Robot Operating System) to read image data and provide a corresponding steering angle in real time. The current model achieves 15.2 degree error on average. As development continues the model may replace the current lane centering software and will be used for IGVC Self-Drive competition and campus transportation.

No abstract available

An approach for modelling adaptive complex systems should be flexible and scalable to allow a system to grow easily, and should have a formal foundation to guarantee the correctness of the system behavior. In this paper, we present the architecture, and formal syntax and semantics of HPobSAM which is a model for specifying behavioral and structural adaptations to model large-scale systems and address re-usability concerns. Self-adaptive modules are used as the building blocks to structure a system, and policies are used as the mechanism to perform both behavioral and structural adaptations. While a self-adaptive module is autonomous to achieve its local goals by collaborating with other self-adaptive modules, it is controlled by a higher-level entity to prevent undesirable behavior. HPobSAM is formalized using a combination of algebraic, graph transformation-based and actor-based formalisms.

No abstract available

No abstract available

No abstract available

No abstract available

No abstract available

Clinical decision-making in oncology is complex, requiring the integration of multimodal data and multidomain expertise. We developed and evaluated an autonomous clinical artificial intelligence (AI) agent leveraging GPT-4 with multimodal precision oncology tools to support personalized clinical decision-making. The system incorporates vision transformers for detecting microsatellite instability and KRAS and BRAF mutations from histopathology slides, MedSAM for radiological image segmentation and web-based search tools such as OncoKB, PubMed and Google. Evaluated on 20 realistic multimodal patient cases, the AI agent autonomously used appropriate tools with 87.5% accuracy, reached correct clinical conclusions in 91.0% of cases and accurately cited relevant oncology guidelines 75.5% of the time. Compared to GPT-4 alone, the integrated AI agent drastically improved decision-making accuracy from 30.3% to 87.2%. These findings demonstrate that integrating language models with precision oncology and search tools substantially enhances clinical accuracy, establishing a robust foundation for deploying AI-driven personalized oncology support systems. Ferber et al. present an autonomous artificial intelligence agent system for deployment of specialized medical oncology computational tools, validating their system across various clinical scenarios representative of typical patient care workflows.

No abstract available

Large language models (LLMs) hold significant promise in the field of medical diagnosis. There are still many challenges in the direct diagnosis of hepatocellular carcinoma (HCC). α-Fetoprotein (AFP) is a commonly used tumor marker for liver cancer. However, relying on AFP can result in missed diagnoses of HCC. We developed an artificial intelligence (AI) agent centered on LLMs, named ChatExosome, which created an interactive and convenient system for clinical spectroscopic analysis and diagnosis. ChatExosome consists of two main components: the first is the deep learning of the Raman fingerprinting of exosomes derived from HCC. Based on a patch-based 1D self-attention mechanism and downsampling, the feature fusion transformer (FFT) was designed to process the Raman spectra of exosomes. It achieved accuracies of 95.8% for cell-derived exosomes and 94.1% for 165 clinical samples, respectively. The second component is the interactive chat agent based on LLM. The retrieval-augmented generation (RAG) method was utilized to enhance the knowledge related to exosomes. Overall, LLM serves as the core of this interactive system, which is capable of identifying users' intentions and invoking the appropriate plugins to process the Raman data of exosomes. This is the first AI agent focusing on exosome spectroscopy and diagnosis, enhancing the interpretability of classification results, enabling physicians to leverage cutting-edge medical research and artificial intelligence techniques to optimize medical decision-making processes, and it shows great potential in intelligent diagnosis.

Evolutionary Game Theory (EGT) and Artificial Intelligence (AI) are two fields that, at first glance, might seem distinct, but they have notable connections and intersections. The former focuses on the evolution of behaviors (or strategies) in a population, where individuals interact with others and update their strategies based on imitation (or social learning). The more successful a strategy is, the more prevalent it becomes over time. The latter, meanwhile, is centered on machine learning algorithms and (deep) neural networks. It is often from a single-agent perspective but increasingly involves multi-agent environments, in which intelligent agents adjust their strategies based on feedback and experience, somewhat akin to the evolutionary process yet distinct in their self-learning capacities. In light of the key components necessary to address real-world problems, including (i) learning and adaptation, (ii) cooperation and competition, (iii) robustness and stability, and altogether (iv) population dynamics of individual agents whose strategies evolve, the cross-fertilization of ideas between both fields will contribute to the advancement of mathematics of multi-agent learning systems, in particular, to the nascent domain of ``collective cooperative intelligence'' bridging evolutionary dynamics and multi-agent reinforcement learning.

We discuss the emerging new opportunity for building feedback‐rich computational models of social systems using generative artificial intelligence. Referred to as generative agent‐based models (GABMs), such individual‐level models utilize large language models to represent human decision‐making in social settings. We provide a GABM case in which human behavior can be incorporated into simulation models by coupling a mechanistic model of human interactions with a pre‐trained large language model. This is achieved by introducing a simple GABM of social norm diffusion in an organization. For educational purposes, the model is intentionally kept simple. We examine a wide range of scenarios and the sensitivity of the results to several changes in the prompt. We hope the article and the model serve as a guide for building useful dynamic models of various social systems that include realistic human reasoning and decision‐making. © 2024 System Dynamics Society.

No abstract available

ChatMOF is an artificial intelligence (AI) system that is built to predict and generate metal-organic frameworks (MOFs). By leveraging a large-scale language model (GPT-4, GPT-3.5-turbo, and GPT-3.5-turbo-16k), ChatMOF extracts key details from textual inputs and delivers appropriate responses, thus eliminating the necessity for rigid and formal structured queries. The system is comprised of three core components (i.e., an agent, a toolkit, and an evaluator) and it forms a robust pipeline that manages a variety of tasks, including data retrieval, property prediction, and structure generations. ChatMOF shows high accuracy rates of 96.9% for searching, 95.7% for predicting, and 87.5% for generating tasks with GPT-4. Additionally, it successfully creates materials with user-desired properties from natural language. The study further explores the merits and constraints of utilizing large language models (LLMs) in combination with database and machine learning in material sciences and showcases its transformative potential for future advancements. LLMs can be augmented with tools to increase their capabilities. Here, authors have developed an artificial intelligence system called ChatMOF combining LLMs and specialised libraries and utilities to predict and generate metal-organic frameworks.

Abstract Intelligent agents must be able to communicate intentions and explain their decision-making processes to build trust, foster confidence, and improve human-agent team dynamics. Recognizing this need, academia and industry are rapidly proposing new ideas, methods, and frameworks to aid in the design of more explainable AI. Yet, there remains no standardized metric or experimental protocol for benchmarking new methods, leaving researchers to rely on their own intuition or ad hoc methods for assessing new concepts. In this work, we present the first comprehensive (n = 286) user study testing a wide range of approaches for explainable machine learning, including feature importance, probability scores, decision trees, counterfactual reasoning, natural language explanations, and case-based reasoning, as well as a baseline condition with no explanations. We provide the first large-scale empirical evidence of the effects of explainability on human-agent teaming. Our results will help to guide the future of explainability research by highlighting the benefits of counterfactual explanations and the shortcomings of confidence scores for explainability. We also propose a novel questionnaire to measure explainability with human participants, inspired by relevant prior work and correlated with human-agent teaming metrics.

Abstract Cloud Computing is one of the most intensively developed solutions for large-scale distributed processing. Effective use of such environments, management of their high complexity and ensuring appropriate levels of Quality of Service (QoS) require advanced monitoring systems. Such monitoring systems have to support the scalability, adaptability and reliability of Cloud. Most of existing monitoring systems do not incorporate any Artificial Intelligence (AI) algorithms for supporting the change inside the task stream or environment itself. They focus only on monitoring or enabling the control of the system as a part of a separated service. An effective monitoring system for the Cloud environment should gather information about all stages of tasks processing and should actively control the monitored environment. In this paper, we present a novel Multi-Agent System based Cloud Monitoring (MAS-CM) model that supports the performance and security of tasks gathering, scheduling and execution processes in large-scale service-oriented environments. Such models are explicitly designed to control the performance and security objectives of the environment. In our work, we focus on prevention of unauthorized task injection and modification, optimization of scheduling process and maximization of resource usage. We evaluate the effectiveness of MAS-CM empirically using an evolutionary driven implementation of Independent Batch Scheduler and FastFlow framework. The obtained results demonstrate the effectiveness of the proposed approach and the performance improvement.

Abstract Applications of wireless sensor networks are blooming for attacking some limits of social development, among which energy consumption and communication latency are fatal. Effective communication traffic control and management is a potential solution, so we propose a novel traffic-control system based on deep reinforcement learning, which regards traffic control as a strategy-learning process, to minimize energy consumption. Our algorithm utilizes deep neural network for learning, inputs the state of wireless sensor network as well as outputs the optimal route path. The simulation experiments demonstrate that our algorithm is feasible to control traffic in wireless sensor network and can reduce the energy consumption.

Traditionally, cognitive and computer scientists have viewed intelligence solipsistically, as a property of unitary agents devoid of social context. Given the success of contemporary learning algorithms, we argue that the bottleneck in artificial intelligence (AI) advancement is shifting from data assimilation to novel data generation. We bring together evidence showing that natural intelligence emerges at multiple scales in networks of interacting agents via collective living, social relationships and major evolutionary transitions, which contribute to novel data generation through mechanisms such as population pressures, arms races, Machiavellian selection, social learning and cumulative culture. Many breakthroughs in AI exploit some of these processes, from multi-agent structures enabling algorithms to master complex games such as Capture-The-Flag and StarCraft II, to strategic communication in the game Diplomacy and the shaping of AI data streams by other AIs. Moving beyond a solipsistic view of agency to integrate these mechanisms could provide a path to human-like compounding innovation through ongoing novel data generation. Advances in machine intelligence often depend on data assimilation, but data generation has been neglected. The authors discuss mechanisms that might achieve continuous novel data generation and the creation of intelligent systems that are capable of human-like innovation, focusing on social aspects of intelligence.

This research aims to demonstrate that artificial intelligence (AI) can function not only as a tool for learning, but also as an intelligent agent with which humans can engage in collaborative learning (CL) to change epistemic practices in science classrooms. We adopted a design and development research approach, following the Analysis, Design, Development, Implementation and Evaluation (ADDIE) model, to prototype a tangible instructional system called Collaborative Learning with AI Speakers (CLAIS). The CLAIS system is designed to have 3–4 human learners join an AI speaker to form a small group, where humans and AI are considered peers participating in the Jigsaw learning process. The development was carried out using the NUGU AI speaker platform. The CLAIS system was successfully implemented in a Science Education course session with 15 pre-service elementary science teachers. The participants evaluated the CLAIS system through mixed methods surveys as teachers, learners, peers, and users. Quantitative data showed that the participants’ Intelligent-Technological, Pedagogical, and Content Knowledge was significantly increased after the CLAIS session, the perception of the CLAIS learning experience was positive, the peer assessment on AI speakers and human peers was different, and the user experience was ambivalent. Qualitative data showed that the participants came to anticipate future changes in the epistemic process in science classrooms, while acknowledging technical issues such as speech recognition and response latency. This study highlights the potential of human-AI collaboration for knowledge co-construction in authentic classroom settings and exemplifies how AI could shape the future landscape of epistemic practices in the classroom.

Exploring rich environments and evaluating one's actions without prior knowledge is immensely challenging. In this paper, we propose Motif, a general method to interface such prior knowledge from a Large Language Model (LLM) with an agent. Motif is based on the idea of grounding LLMs for decision-making without requiring them to interact with the environment: it elicits preferences from an LLM over pairs of captions to construct an intrinsic reward, which is then used to train agents with reinforcement learning. We evaluate Motif's performance and behavior on the challenging, open-ended and procedurally-generated NetHack game. Surprisingly, by only learning to maximize its intrinsic reward, Motif achieves a higher game score than an algorithm directly trained to maximize the score itself. When combining Motif's intrinsic reward with the environment reward, our method significantly outperforms existing approaches and makes progress on tasks where no advancements have ever been made without demonstrations. Finally, we show that Motif mostly generates intuitive human-aligned behaviors which can be steered easily through prompt modifications, while scaling well with the LLM size and the amount of information given in the prompt.

Significance Humans have long demonstrated an ability to learn from interactions with others. However, artificial intelligence (AI) agents learn in social isolation. To create intelligent systems that understand more than a fixed slice of the world, our article formalizes socially situated AI—a framework that enables agents to interact with people as they simultaneously learn new concepts about the world around them. Using our framework, we deploy a field experiment on a photo-sharing social network where our agent interacts with hundreds of thousands of people to learn concepts about the visual world. We combine advances in deep learning, computer vision, natural language processing, and human–computer Interaction to deliver a human-centered AI that learns from interactions with people in social environments.

We introduce MAgent, a platform to support research and development of many-agent reinforcement learning. Unlike previous research platforms on single or multi-agent reinforcement learning, MAgent focuses on supporting the tasks and the applications that require hundreds to millions of agents. Within the interactions among a population of agents, it enables not only the study of learning algorithms for agents' optimal polices, but more importantly, the observation and understanding of individual agent's behaviors and social phenomena emerging from the AI society, including communication languages, leaderships, altruism. MAgent is highly scalable and can host up to one million agents on a single GPU server. MAgent also provides flexible configurations for AI researchers to design their customized environments and agents. In this demo, we present three environments designed on MAgent and show emerged collective intelligence by learning from scratch.

High-performing human teams leverage intelligent and efficient communication and coordination strategies to collaboratively maximize their joint utility. Inspired by teaming behaviors among humans, I seek to develop computational methods for synthesizing intelligent communication and coordination strategies for collaborative multi-robot systems. I leverage both classical model-based control and planning approaches as well as data-driven methods such as Multi-Agent Reinforcement Learning (MARL) to provide several contributions towards enabling emergent cooperative teaming behavior across both homogeneous and heterogeneous (including agents with different capabilities) robot teams.

This article focuses on the research of artificial intelligence agents based on large language models. These agents break away from the traditional reinforcement learning framework and can achieve internal-driven evolution through their own language generation. The article details several representative research results, including the HPTSA system from the University of Illinois at Urbana-Champaign, which adopts a hierarchical planning and task-specific agent collaboration model and significantly improves efficiency in zero-day vulnerability attacks, outperforming single-agent systems and open-source vulnerability scanners; the BattleAgent multimodal dynamic simulation system from Rutgers University, which can simulate the complex dynamic interactions of agents and provide support for historical battle reenactments; the WarAgent multi-agent simulation system from Rutgers and the University of Michigan, which can simulate international conflict events to explore factors related to war and peace; the general embodied intelligent agent research and series projects from NVIDIA, which promote the development of embodied intelligence; the "Unified Agent" framework from DeepMind of Google, which alleviates some drawbacks of traditional reinforcement learning technology; and the "Smallville" platform from Stanford University and Google's Artificial Intelligence Research Institute, as well as the Dynalang intelligent agent from the University of California, Los Angeles. These studies demonstrate the powerful capabilities and wide application prospects of artificial intelligence agents empowered by large language models in various fields.

Delivering intelligent and adaptive navigation assistance in augmented reality (AR) requires more than visual cues, as it demands systems capable of interpreting flexible user intent and reasoning over both spatial and semantic context. Prior AR navigation systems often rely on rigid input schemes or predefined commands, which limit the utility of rich building data and hinder natural interaction. In this work, we propose an embodied AR navigation system that integrates Building Information Modeling (BIM) with a multi-agent retrieval-augmented generation (RAG) framework to support flexible, language-driven goal retrieval and route planning. The system orchestrates three language agents, Triage, Search, and Response, built on large language models (LLMs), which enables robust interpretation of open-ended queries and spatial reasoning using BIM data. Navigation guidance is delivered through an embodied AR agent, equipped with voice interaction and locomotion, to enhance user experience. A real-world user study yields a System Usability Scale (SUS) score of 80.5, indicating excellent usability, and comparative evaluations show that the embodied interface can significantly improves users' perception of system intelligence. These results underscore the importance and potential of language-grounded reasoning and embodiment in the design of user-centered AR navigation systems. Video demonstrations are available at https://woven-visionai.github.io/ar-navigation-agent-project.

In intelligent transportation systems, the convergence of large language models and embodied artificial intelligence (AI) has led to Vehicular Embodied AI Networks (VEANs). In VEANs, the integration of digital twins results in Vehicular Embodied Agent AI Twins (VEAATs), which enables AI assistants for in-vehicle application services. Constrained by computational latency and limited on-board computing resources, autonomous vehicles (AVs) must offload computationally intensive VEAATs to Roadside Units (RSUs) for remote execution support with substantial resources. Considering the limited coverage of RSUs and the mobility of AVs, the migration of VEAATs among RSUs is crucial for ensuring continuous and high-quality immersive services. Nevertheless, high-density traffic exacerbates the workload imbalance among RSUs, increasing the risk of overload. This paper models the interaction between AVs and RSUs as a Stackelberg game to optimize bandwidth resource allocation for efficient task migration. We employ a multi-agent deep reinforcement learning algorithm to approximate the Stackelberg Equilibrium (SE) and address the workload imbalance issue. Specifically, we propose Path eXclusion-based Multi-Agent Proximal Policy Optimization (PX-MAPPO), which integrates the PX-based neural network pruning algorithm with the MAPPO algorithm. Numerical results demonstrate that this algorithm reduces the number of network parameters, decreases model complexity, and minimizes performance loss.

Much of our everyday, embodied action comes in the form of smooth coping. Smooth coping is skillful action that has become habituated and ingrained, generally placing less stress on cognitive load than considered and deliberative thought and action. When performed with skill and expertise, walking, driving, skiing, musical performances, and short-order cooking are all examples of the phenomenon. Smooth coping is characterized by its rapidity and relative lack of reflection, both being hallmarks of automatization. Deliberative and reflective actions provide the contrast case. In Dreyfus’ classic view, smooth coping is “mindless” absorption into action, being in the flow, and any reflective thought will only interrupt this flow. Building on the pragmatist account of Dewey, others, such as Sutton, Montero, and Gallagher, insist on the intelligent flexibility built into smooth coping, suggesting that it is not equivalent to automatization. We seek to answer two complementary challenges in this article. First, how might we model smooth coping in autonomous agents (natural or artificial) at fine granularity? Second, we use this model of smooth coping to show how we might implement smooth coping in artificial intelligent agents. We develop a conceptual model of smooth coping in LIDA (Learning Intelligent Decision Agent). LIDA is an embodied cognitive architecture implementing the global workspace theory of consciousness, among other psychological theories. LIDA’s implementation of consciousness enables us to account for the phenomenology of smooth coping, something that few cognitive architectures would be able to do. Through the fine granular analysis of LIDA, we argue that smooth coping is a sequence of automatized actions intermittently interspersed with consciously mediated action selection, supplemented by dorsal stream processes. In other words, non-conscious, automatized actions (whether learned or innate) often require occasional bursts of conscious cognition to achieve the skillful and flexible adjustments of smooth coping. In addition, never-conscious dorsal stream information and associated sensorimotor processes provide further online adjustments during smooth coping. To achieve smooth coping in LIDA we introduce a new module to the LIDA cognitive architecture the Automatized Action Selection sub-module. Our complex model of smooth coping borrows notions of “embodied intelligence” from enactivism and augments these by allowing representations and more detailed mechanisms of conscious control. We explore several extended examples of smooth coping, starting from basic activities like walking and scaling up to more complex tasks like driving and short-order cooking.

Intelligent assistants powered by large-language models (LLMs) are increasingly becoming a part of our daily lives, particularly on our mobile devices. In immersive virtual reality (VR), intelligent assistants can be embodied as intelligent virtual agents (IVAs). They are commonly humanoid characters capable of multimodal interaction with the user. When prompted with a question, these assistants or IVAs typically formulate elaborate responses, which are then vocalized. While it is possible to terminate the output by tapping on a mobile phone screen, in immersive VR, different strategies have to be investigated to interrupt the IVA. In our work, we investigated whether implementing the possibility to interrupt embodied IVAs in VR enhances the user experience and the users’ perception of the IVA. Therefore, we conducted a user study in which 30 participants were tasked to answer several quiz questions with the help of an IVA, with or without the ability to interrupt. The results show that the possibility of interrupting the IVA increases perceived efficiency, attractiveness and stimulation. Qualitative feedback suggests that real-time interruptions lead to more natural and dynamic conversations. Although the interruption possibility did not change users’ perception of the IVA regarding perceived anthropomorphism, likeability, and perceived intelligence, we discussed these findings along with participants’ qualitative feedback and provide various design insights for future research.

Natural language serves as the primary mode of communication when an intelligent agent with a physical presence engages with human beings. While a plethora of research focuses on natural language understanding (NLU), encompassing endeavors such as sentiment analysis, intent prediction, question answering, and summarization, the scope of NLU directed at situations necessitating tangible actions by an embodied agent remains limited. The inherent ambiguity and incompleteness inherent in natural language present challenges for intelligent agents striving to decipher human intention. To tackle this predicament head-on, we introduce a novel system known as task and argument grounding for Embodied agents (tagE). At its core, our system employs an inventive neural network model designed to extract a series of tasks from complex task instructions expressed in natural language. Our proposed model adopts an encoder-decoder framework enriched with nested decoding to effectively extract tasks and their corresponding arguments from these intricate instructions. These extracted tasks are then mapped (or grounded) to the robot's established collection of skills, while the arguments find grounding in objects present within the environment. To facilitate the training and evaluation of our system, we have curated a dataset featuring complex instructions. The results of our experiments underscore the prowess of our approach, as it outperforms robust baseline models.

Interactive embodied agents are intelligent agents present inside some virtual environment and interact with humans to do tasks we want them to simulate. The communication takes place in some verbal or non-verbal form. This kind of communication gives a better and richer experience and more possibility for users to be satisfied with the carried-out task. An intelligent agent analyzes the environment it is in and performs tasks with all its knowledge based on the interaction. The task is carried out in the game environment where the Agent resides. There are some challenges to an agent that it needs to address, as well as to understand the instruction given to it. The next challenge is asking for clarification if the instruction given is unclear. In this research paper, we have discussed the problems mentioned above using the dataset provided by the NeurIPS 2022 IGLU Challenge which is a labeled dataset. We have discussed the various approaches for the problem mentioned.

Recent advances in embodied agents with multimodal perception and reasoning capabilities based on large vision-language models (LVLMs), excel in autonomously interacting either real or cyber worlds, helping people make intelligent decisions in complex environments. However, the current works are normally optimized by golden action trajectories or ideal task-oriented solutions toward a definitive goal. This paradigm considers limited user-oriented factors, which could be the reason for their performance reduction in a wide range of personal assistant applications. To address this, we propose Chain-of-User-Thought (COUT, a novel embodied reasoning paradigm that takes a chain of thought from basic action thinking to explicit and implicit personalized preference thought to incorporate personalized factors into autonomous agent learning. The main challenges of achieving COUT include: 1) the definition of embodied personalized tasks, 2) the embodied environment epitomizes personalized preference, and 3) the way to model embodied personalized actions. To target COUT, we introduce SmartAgent, an agent framework perceiving cyber environments and reasoning personalized requirements as: 1) interacting with GUI to access an item pool, 2) generating users' explicit requirements implied by previous actions, and 3) recommending items to fulfill users' implicit requirements. To demonstrate SmartAgent's capabilities, we also create a brand-new dataset SmartSpot that offers a full-stage personalized action-involved environment. To our best knowledge, our work is the first to formulate the COUT process, serving as a preliminary attempt towards embodied personalized agent learning. Our extensive experiments on SmartSpot illuminate SmartAgent’s functionality among a series of embodied and personalized sub-tasks.

The integration of conversational agents in virtual reality (VR) offers promising opportunities for applications such as training, education, and mental health support. However, most current VR agents lack empathic capabilities, which are crucial for fostering engagement, trust, and effective interaction. To address this gap, we present EMBRACE, a modular framework that combines natural language processing, emotion inference, expressive text-to-speech synthesis, and AI-driven facial animation to deliver emotionally responsive VR agents. We implemented EMBRACE within a counselling-inspired virtual environment and conducted a user study, comparing empathic and nonempathic agent conditions. Subjective measures included usability, presence, and agent evaluation, while objective measures incorporated heart rate and heart rate variability. Results show that empathic behaviours enhanced spatial presence and revealed trends toward higher general presence and physiological engagement. Additional analyses indicated that prior chatbot experience shaped usability and agent ratings, while gender influenced certain presence dimensions. Our findings demonstrate both the feasibility and benefits of implementing empathic VR agents.

No abstract available

No abstract available

Embodied conversational agents have changed the ways we can interact with machines. However, these systems often do not meet users’ expectations. A limitation is that the agents are monotonic in behavior and do not adapt to an interlocutor. We present SIVA (a Socially Intelligent Virtual Agent), an expressive, embodied conversational agent that can recognize human behavior during open-ended conversations and automatically align its responses to the conversational and expressive style of the other party. SIVA leverages multimodal inputs to produce rich and perceptually valid responses (lip syncing and facial expressions) during the conversation. We conducted a user study (N=30) in which participants rated SIVA as being more empathetic and believable than the control (agent without style matching). Based on almost 10 hours of interaction, participants who preferred interpersonal involvement evaluated SIVA as significantly more animate than the participants who valued consideration and independence.

No abstract available

The research and development (R&D) of intelligent virtual agents (IVAs) is inherently complex. We aim to manage this complexity by combining the best aspects of academic and commercial approaches into a principled R&D platform that emphasizes interoperability, ex-tendability, re-use, and support for multiple hardware targets. This IVA platform, the Virtual Human Toolkit 2.0, is a re-architecture of our earlier work and combines a modular message passing architecture with that of a microservices architecture. This paper discusses our approach, design decisions, lessons learned, and current status of this ongoing effort. We illustrate the strengths of the architecture, how best to use commodity AI cloud services in one's own work, and how to port legacy stand-alone software to a web service.

No abstract available

Embodied Artificial Intelligence (AI) bridges the cyberspace and the physical space, driving advancements in autonomous systems like the Vehicular Embodied AI NETwork (VEANET). VEANET integrates advanced AI capabilities into vehicular systems to enhance autonomous operations and decision-making. Embodied agents, such as Autonomous Vehicles (AVs), are autonomous entities that can perceive their environment and take actions to achieve specific goals, actively interacting with the physical world. Embodied Agent Twins (EATs) are digital models of these embodied agents, with various Embodied Agent AI Twins (EAATs) for intelligent applications in cyberspace. In VEANETs, EAATs act as in-vehicle AI assistants to perform diverse tasks supporting autonomous driving using generative AI models. Due to limited onboard computational resources, AVs offload EAATs to nearby RoadSide Units (RSUs). However, the mobility of AVs and limited RSU coverage necessitates dynamic migrations of EAATs, posing challenges in selecting suitable RSUs under information asymmetry. To address this, we construct a multi-dimensional contract theoretical model between AVs and alternative RSUs. Considering that AVs may exhibit irrational behavior, we utilize prospect theory instead of expected utility theory to model the actual utilities of AVs. Finally, we employ a Generative Diffusion Model (GDM)-based algorithm to identify the optimal contract designs, thus enhancing the efficiency of EAAT migrations. Numerical results demonstrate the superior efficiency of the proposed GDM-based scheme in facilitating EAAT migrations compared with traditional deep reinforcement learning methods.

This paper presents a computational model for managing an Embodied Conversational Agent's first impressions of warmth and competence towards the user. These impressions are important to manage because they can impact users' perception of the agent and their willingness to continue the interaction with the agent. The model aims at detecting user's impression of the agent and producing appropriate agent's verbal and nonverbal behaviours in order to maintain a positive impression of warmth and competence. User's impressions are recognized using a machine learning approach with facial expressions (action units) which are important indicators of users' affective states and intentions. The agent adapts in real-time its verbal and nonverbal behaviour, with a reinforcement learning algorithm that takes user's impressions as reward to select the most appropriate combination of verbal and non-verbal behaviour to perform. A user study to test the model in a contextualized interaction with users is also presented. Our hypotheses are that users' ratings differs when the agents adapts its behaviour according to our reinforcement learning algorithm, compared to when the agent does not adapt its behaviour to user's reactions (i.e., when it randomly selects its behaviours). The study shows a general tendency for the agent to perform better when using our model than in the random condition. Significant results shows that user's ratings about agent's warmth are influenced by their a-priori about virtual characters, as well as that users' judged the agent as more competent when it adapted its behaviour compared to random condition.

Intelligent virtual agents (VAs) already support us in a variety of everyday tasks such as setting up appointments, monitoring our fitness, and organizing messages. Adding a humanoid body representation to these mostly voice-based VAs has enormous potential to enrich the human–agent communication process but, at the same time, raises expectations regarding the agent’s social, spatial, and intelligent behavior. Embodied VAs may be perceived as less human-like if they, for example, do not return eye contact, or do not show a plausible collision behavior with the physical surroundings. In this article, we introduce a new model that extends human-to-human interaction to interaction with intelligent agents and covers different multi-modal and multi-sensory channels that are required to create believable embodied VAs. Theoretical considerations of the different aspects of human–agent interaction are complemented by implementation guidelines to support the practical development of such agents. In this context, we particularly emphasize one aspect that is distinctive of embodied agents, i.e., interaction with the physical world. Since previous studies indicated negative effects of implausible physical behavior of VAs, we were interested in the initial responses of users when interacting with a VA with virtual–physical capabilities for the first time. We conducted a pilot study to collect subjective feedback regarding two forms of virtual–physical interactions. Both were designed and implemented in preparation of the user study, and represent two different approaches to virtual–physical manipulations: (i) displacement of a robotic object, and (ii) writing on a physical sheet of paper with thermochromic ink. The qualitative results of the study indicate positive effects of agents with virtual–physical capabilities in terms of their perceived realism as well as evoked emotional responses of the users. We conclude with an outlook on possible future developments of different aspects of human–agent interaction in general and the physical simulation in particular.

Non-adherence to a treatment plan recommended by the therapist is a key cause of the increasing rate of chronic medical conditions globally. The therapist-patient therapeutic alliance is regarded as a successful intervention and a good predictor of treatment adherence. Similar to the human scenario, embodied conversational agents (ECAs) showed evidence of their ability to build an agent-patient therapeutic alliance, which motivates the effort to advance ECAs as a potential solution to improve treatment adherence and consequently the health outcome. Building therapeutic alliance implies the need for a positive environment where the ECA and the patient can share their knowledge and discuss their goals, preferences and tasks towards building a shared plan, which is commonly done using explanations. However, explainable agents commonly rely on their own knowledge and goals in providing explanations, rather than the beliefs, plans or goals of the user. It is not clear whether such explanations, in individual-specific contexts such as personal health assistance, are perceived by the user as relevant in decision-making towards their own behavior change. Therefore, in this research, we are developing a user-aware explainable ECA by embedding the cognitive agent architecture with a user model, explanation engine and modified planner to implement the concept of SharedPlans. The developed agent will be deployed and evaluated with real patients and the therapeutic alliance will be measured using standard measurements.

Video creation has become increasingly popular, yet the expertise and effort required for editing often pose barriers to beginners. In this paper, we explore the integration of large language models (LLMs) into the video editing workflow to reduce these barriers. Our design vision is embodied in LAVE, a novel system that provides LLM-powered agent assistance and language-augmented editing features. LAVE automatically generates language descriptions for the user’s footage, serving as the foundation for enabling the LLM to process videos and assist in editing tasks. When the user provides editing objectives, the agent plans and executes relevant actions to fulfill them. Moreover, LAVE allows users to edit videos through either the agent or direct UI manipulation, providing flexibility and enabling manual refinement of agent actions. Our user study, which included eight participants ranging from novices to proficient editors, demonstrated LAVE’s effectiveness. The results also shed light on user perceptions of the proposed LLM-assisted editing paradigm and its impact on users’ creativity and sense of co-creation. Based on these findings, we propose design implications to inform the future development of agent-assisted content editing.

Recent advancements in foundation models (FMs) have unlocked new prospects in autonomous driving, yet the experimental settings of these studies are preliminary, oversimplified, and fail to capture the complexity of real-world driving scenarios in human environments. It remains under-explored whether FM agents can handle long-horizon navigation tasks with free-from dialogue and deal with unexpected situations caused by environmental dynamics or task changes. To explore the capabilities and boundaries of FMs faced with the challenges above, we introduce DriVLMe, a video-language-model-based agent to facilitate natural and effective communication between humans and autonomous vehicles that perceive the environment and navigate. We develop DriVLMe from both embodied experiences in a simulated environment and social experiences from real human dialogue. While DriVLMe demonstrates competitive performance in both open-loop benchmarks and closed-loop human studies, we reveal several limitations and challenges, including unacceptable inference time, imbalanced training data, limited visual understanding, challenges with multi-turn interactions, simplified language generation from robotic experiences, and difficulties in handling on-the-fly unexpected situations like environmental dynamics and task changes. Nevertheless, DriVLMe offers a promising new direction for autonomous driving agents that need to navigate not just complex environments but also complex social interactions.

Building natural and conversational virtual humans is a task of formidable complexity. We believe that, especially when building agents that affectively interact with biological humans in real-time, a cognitive science-based, multilayered sensing and artificial intelligence (AI) systems approach is needed. For this demo, we show a working version (through human interaction with it) our modular system of natural, conversation 3D virtual human using AI or sensing layers. These including sensing the human user via facial emotion recognition, voice stress, semantic meaning of the words, eye gaze, heart rate, and galvanic skin response. These inputs are combined with AI sensing and recognition of the environment using deep learning natural language captioning or dense captioning. These are all processed by our AI avatar system allowing for an affective and empathetic conversation using an NLP topic-based dialogue capable of using facial expressions, gestures, breath, eye gaze and voice language-based two-way back and forth conversations with a sensed human. Our lab has been building these systems in stages over the years.

Embodied intelligence empowers agents with a profound sense of perception, enabling them to respond in a manner closely aligned with real-world situations. Large Language Models (LLMs) delve into language instructions with depth, serving a crucial role in generating plans for intricate tasks. Thus, LLM-based embodied models further enhance the agent's capacity to comprehend and process information. However, this amalgamation also ushers in new challenges in the pursuit of heightened intelligence. Specifically, attackers can manipulate LLMs to produce irrelevant or even malicious outputs by altering their prompts. Confronted with this challenge, we observe a notable absence of multi-modal datasets essential for comprehensively evaluating the robustness of LLM-based embodied models. Consequently, we construct the Embodied Intelligent Robot Attack Dataset (EIRAD), tailored specifically for robustness evaluation. Additionally, two attack strategies are devised, including untargeted attacks and targeted attacks, to effectively simulate a range of diverse attack scenarios. At the same time, during the attack process, to more accurately ascertain whether our method is successful in attacking the LLM-based embodied model, we devise a new attack success evaluation method utilizing the BLIP2 model. Recognizing the time and cost-intensive nature of the GCG algorithm in attacks, we devise a scheme for prompt suffix initialization based on various target tasks, thus expediting the convergence process. Experimental results demonstrate that our method exhibits a superior attack success rate when targeting LLM-based embodied models, indicating a lower level of decision-level robustness in these models.

In this work, we introduce SMART-LLM, an innovative framework designed for embodied multi-robot task planning. SMART-LLM: Smart Multi-Agent Robot Task Planning using Large Language Models (LLMs), harnesses the power of LLMs to convert high-level task instructions provided as input into a multi-robot task plan. It accomplishes this by executing a series of stages, including task decomposition, coalition formation, and task allocation, all guided by programmatic LLM prompts within the few-shot prompting paradigm. We create a benchmark dataset designed for validating the multi-robot task planning problem, encompassing four distinct categories of high-level instructions that vary in task complexity. Our evaluation experiments span both simulation and real-world scenarios, demonstrating that the proposed model can achieve promising results for generating multi-robot task plans. The experimental videos, code, and datasets from the work can be found at https://sites.google.com/view/smart-llm/.

The promising potential of AI and network convergence in improving networking performance and enabling new service capabilities has recently attracted significant interest. Existing network AI solutions, while powerful, are mainly built based on the close-loop and passive learning framework, resulting in major limitations in autonomous solution finding and dynamic environmental adaptation. Agentic AI has recently been introduced as a promising solution to address the above limitations and pave the way for true, generally intelligent, and beneficial AI systems. The key idea is to create a networking ecosystem to support a diverse range of autonomous and embodied AI agents in fulfilling their goals. In this article, we focus on the novel challenges and requirements of agentic AI networking. We propose AgentNet, a novel framework for supporting interaction, collaborative learning, and knowledge transfer among AI agents. We introduce a general architectural framework of AgentNet and then propose a generative foundation model (GFM)-based implementation in which multiple GFM-as-agents have been created as an interactive knowledge-base to bootstrap the development of embodied AI agents according to different task requirements and environmental features. We consider two application scenarios, digital-twin-based industrial automation and metaverse-based infotainment system, to describe how to apply AgentNet for supporting efficient task-driven collaboration and interaction among AI agents.

No abstract available

Our research investigates the impact of latency on presence and immersion in virtual reality (VR) environments, focusing on interactions with LLM-powered Embodied Conversational Agents (ECAs). We explore the effectiveness of multimodal feedback strategies—including filled pauses, nonverbal turn-taking behaviours, and visual feedback—in mitigating perceived latency. Eighteen participants were subjected to both a baseline condition, without feedback interventions, and a feedback-enhanced condition. Our findings indicate that the feedback condition significantly improved the sense of presence and immersion. We also found that perceived response time and users’ impressions of the agents improved, thereby increasing willingness for future interactions. Additionally, chatbot experience positively correlated with agent likeability, whereas VR experience showed no significant correlation. These results highlight the effectiveness of feedback modalities in enhancing spatial presence and overall immersion, despite latency issues in VR interactions with LLM-powered agents.

No abstract available

: Microsoft Entra ID is Microsoft’s identity and access management solution used by many public and private sector organisations globally. In March 2023, Microsoft retired two PowerShell modules which have enabled automation of administrative tasks, such as user management. The replacement module is based on Microsoft Graph API, and its effective usage would require administrators to learn software development skills. In this paper, we will report the results of work-in-progress research on exploring the applicability of LLM-powered autonomous agents to solve real-life problems. We describe the design and proof-of-concept implementation of MEAN, an agent that performs Entra ID administrative tasks using Microsoft Graph API based on natural language prompts. The results show that LLM-powered autonomous agents can perform at least simple Entra ID administrative tasks. This indicates that the agents could ease the administrative burden by removing the need to learn software development skills.

Cloud Operations (CloudOps) is a rapidly growing field focused on the automated management and optimization of cloud infrastructure which is essential for organizations nav-igating increasingly complex cloud environments. MontyCloud Inc. is one of the major companies in the CloudOps domain that leverages autonomous bots to manage cloud compliance, security, and continuous operations. To make the platform more accessible and effective to the customers, we leveraged the use of GenAl. Developing a GenAl-based solution for autonomous CloudOps for the existing MontyCloud system presented us with various challenges such as i) diverse data sources; ii) orchestration of multiple processes and iii) handling complex workflows to automate routine tasks. To this end, we developed MOYA, a multi-agent framework that leverages GenAI and balances autonomy with the necessary human control. This framework integrates various internal and external systems and is optimized for factors like task orchestration, security, and error mitigation while producing accurate, reliable, and relevant insights by utilizing Retrieval Augmented Generation (RAG). Evaluations of our multi-agent system with the help of practitioners as well as using automated checks demonstrate enhanced accuracy, responsiveness, and effectiveness over non-agentic approaches across complex workflows.

This paper explores the integration of symbolic and connectionist paradigms within the realm of Large Language Model (LLM)-powered autonomous agents, highlighting the complementary strengths of each approach. Symbolic AI, known for its structured, rule-based logic, excels at encoding explicit knowledge and facilitating reasoning, while connectionist AI, particularly neural networks, provides robustness in handling large-scale unstructured data through learning from examples. By merging these paradigms, we propose a synergistic framework that enhances autonomous agent capabilities in both reasoning and adaptability. We investigate how LLMs, which exhibit traits of both paradigms, can serve as the backbone for this integration, fostering improved decision-making, natural language understanding, and autonomy. Our findings underscore the potential of this hybrid approach to advance the development of intelligent agents that can navigate complex environments, reason effectively, and learn from experience in dynamic, real-world applications.

We demonstrate the first cross-domain cross-layer level-4 autonomous optical network via a multi-AI-agent system. Field trials show ~98% task completion rate across the distributed AI training lifecycle©3.2© higher than single agents using advanced LLMs. ©2025 The Author(s)

: Infrastructure failures like the 2024 CrowdStrike incident in July demonstrate the critical need for autonomous recovery systems that can detect, diagnose, and remediate outages without the dependency on human intervention. We present Sura.ai, a multi-agent system using four cooperating Fetch.ai uAgents orchestrated via the Agentverse Mailbox to provide autonomous infrastructure recovery. Our system integrates LLM-powered root cause analysis through Claude Sonnet 3.5, enabling intelligent decision-making beyond simple rule-based automation. We conducted comprehensive testing across four disaster scenarios including simulated faulty updates, CPU spikes, and cascading failures. Sura.ai achieved a 97.6% action success rate (41/42 incidents resolved) with intelligent alert deduplication across all tested scenarios. Our work demonstrates the feasibility of LLM-based multi-agent orchestration for critical infrastructure resilience and introduces practical patterns for AgentOps implementation in the growing agentic cybersecurity space.

We design and demonstrate the first field trial of LLM-powered AI Agent for ADON. Three operation modes of the Agent are proposed for network lifecycle management to process wavelength add/drop, soft/hard failures, and power optimizations.

The integration of a complex set of electronic design automation (EDA) tools to enhance interoperability is a critical concern for circuit designers. Recent advancements in large language models (LLMs) have showcased their exceptional capabilities in natural language processing and comprehension, offering a novel approach to interfacing with EDA tools. This research article introduces ChatEDA, an autonomous agent for EDA empowered by an LLM, AutoMage, complemented by EDA tools serving as executors. ChatEDA streamlines the design flow from the register-transfer level (RTL) to the graphic data system version II (GDSII) by effectively managing task decomposition, script generation, and task execution. Through comprehensive experimental evaluations, ChatEDA has demonstrated its proficiency in handling diverse requirements, and our fine-tuned AutoMage model has exhibited superior performance compared to GPT-4 and other similar LLMs.

We demonstrate that a group of Al agents can autonomously optimize power commissioning in WDM links. By leveraging modern LLMs' reflection and reasoning capabilities and interacting with a network digital twin, the agents achieve optimal solutions for different criteria such as power and OSNR equalization.

: Large language models (LLMs) have revolutionized the field of artificial intelligence, endowing it with sophisticated language understanding and generation capabilities. However, when faced with more complex and interconnected tasks that demand a profound and iterative thought process, LLMs reveal their inherent limitations. Autonomous LLM-powered multi-agent systems represent a strategic response to these challenges. While these architectures hold promising potential in amplifying AI capabilities, striking the right balance be-tween different levels of autonomy and alignment remains the crucial challenge for their effective operation. This paper proposes a comprehensive multi-dimensional taxonomy, engineered to analyze how autonomous LLM-powered multi-agent systems balance the dynamic interplay between autonomy and alignment across various aspects inherent to architectural viewpoints such as goal-driven task management, agent composition, multi-agent collaboration, and context interaction. Our taxonomy aims to empower researchers, engineers, and AI practitioners to systematically analyze the architectural dynamics and balancing strategies employed by these increasingly prevalent AI systems. The exploratory taxonomic classification of selected representative LLM-powered multi-agent systems illustrates its practical utility and reveals potential for future research and development. An extended version of this paper is available on arXiv (Händler, 2023).

No abstract available

Large language models (LLMs) have catalyzed artificial intelligence (AI) agent development for autonomous optical networks (AONs). This invited paper reviews our recent progress in leveraging LLM-powered AI agents for realizing AONs in field-deployed networks.

The unique integration of Large Language Models (LLMs) as the reasoning center inside Agentic Artificial Intelligence (AAI) systems exposes new, systematic hazards that current research is totally ill-fitted to handle. The lack of a consistent, thorough framework that simultaneously addresses the basic problems of LLM unreliability as they spread and grow in autonomous, goal-directed, multi-agent systems reveals a major gap in the literature. These issues reach important, under-investigated fields like avoiding goal misalignment, lowering the danger of opaque decision-making, and guaranteeing strong long-term safety in complicated systems. Clearly establishing moral and legal responsibility and existing means for minimal human supervision are clearly lacking, therefore creating a hazardous hole as these automated systems approach actual deployment. This article provides an innovative, unified framework for responsible development, thus directly tackling this critical need. Specifically intended for LLM-powered agentic systems operating in challenging, high-stakes contexts, the author has presented the Trust, Risk, and Safety Management (TRiSM) governance framework. The core innovation of the framework is the Goal-Constraint Alignment (GCA) mechanism, which dynamically monitors and constrains LLM behavior inside set ethical and safety envelopes, hence acting as a dynamic barrier against both planned and unexpected goal misalignment. Furthermore, we install a Decentralized Oversight Ledger (DOL) to improve transparency and allow realistic accountability. The DOL offers real-time, tamper-proof, auditable tracking of all multi-agent interactions and decisions, therefore enhancing human oversight and establishing a clear chain of custody for agent behavior, which is vital to determine legal responsibility. Studies show a major improvement in systematic safety and a significant decrease in catastrophic failures compared to present baseline systems by verifying the effectiveness of the framework against a fresh collection of high-stakes, multi-agent coordination scenarios. This study offers the essential technical and governance structure needed for the responsible and safe distribution of next-generation autonomous artificial intelligence.

Recently, LLM-powered driver agents have demonstrated considerable potential in the field of autonomous driving, showcasing human-like reasoning and decision-making abilities. However, current research on aligning driver agent behaviors with human driving styles remains limited, partly due to the scarcity of high-quality natural language data from human driving behaviors. To address this research gap, we propose a multi-alignment framework designed to align driver agents with human driving styles through demonstrations and feedback. Notably, we construct a natural language dataset of human driver behaviors through naturalistic driving experiments and post-driving interviews, offering high-quality human demonstrations for LLM alignment. The framework’s effectiveness is validated through simulation experiments in the CARLA urban traffic simulator and further corroborated by human evaluations. Our research offers valuable insights into designing driving agents with diverse driving styles. The implementation of the framework 1 and details of the dataset 2 can be found at the link.

In recent years, recommendation systems have evolved from providing a single list of recommendations to offering a comprehensive suite of topic-focused services. To better accomplish this task, conversational recommendation systems (CRS) have progressed from basic retrieval-augmented LLM generation to agentic systems with advanced reasoning and self-correction capabilities. However, agentic systems come with notable response latency—a longstanding challenge for conversational recommendation systems. To balance the trade-off between handling complex queries and minimizing latency, we propose AdaptJobRec, the first conversational job recommendation system that leverages autonomous agent to integrate personalized recommendation algorithm tools. The system employs a user query complexity identification mechanism to minimize response latency. For straightforward queries, the agent directly selects the appropriate tool for rapid responses. For complex queries, the agent uses the memory processing module to filter chat history for relevant content, then passes the results to the intelligent task decomposition planner, and finally executes the tasks using personalized recommendation tools. Evaluation on Walmart’s real-world career recommendation scenarios demonstrates that AdaptJobRec reduces average response latency by up to 53.3\% compared to competitive baselines, while significantly improving recommendation accuracy.

Text-based applications and chatbots are increasingly popular for delivering banking services and educational tools, offering convenient and efficient solutions for users. Whereas, personalized assistants have transformed user engagement in the digital banking space by utilizing Large Language Models (LLMs) in conjunction with autonomous agents. This study proposes the development of an intelligent personalized assistant for digital banking, utilizing a multi-agent framework based on the LangGraph and Chain of Thoughts (COT) prompting. While COT guarantees context-aware replies, the LangGraph design maps characteristics to nodes to improve user interactions. The objectives of this system are to enhance task efficiency and elevate the capabilities of digital banking assistants. We present a customizable digital banking system powered by LLM-based models, designed to deliver an interactive and personalized banking experience. The system supports a range of services, including adding money, transferring funds, paying bills, accessing telco services like mobile recharge, managing savings interest rates, DPS schemes, fixed deposits, and answering FAQs related to banking information. Therefore, integrating COT for logical reasoning enhances the effectiveness of multi-agent systems, as each single agent benefits from the structured reasoning process. In addition, LangGraph is employed for structured data management, enabling the assistant to support and accelerate various digital banking processes efficiently. The code implementation of this work is available for public access at: https://github.com/srv-sh/digital_agent.

Current service robots suffer from limited natural language communication abilities, heavy reliance on predefined commands, ongoing human intervention, and, most notably, a lack of proactive collaboration awareness in human-populated environments. This results in narrow applicability and low utility. In this paper, we introduce AssistantX, an LLM-powered proactive assistant designed for autonomous operation in real-world scenarios with high accuracy. AssistantX employs a multi-agent framework consisting of 4 specialized LLM agents, each dedicated to perception, planning, decision-making, and reflective review, facilitating advanced inference capabilities and comprehensive collaboration awareness, much like a human assistant by your side. We built a dataset of 210 real-world tasks to validate AssistantX, which includes instruction content and status information on whether relevant personnel are available. Extensive experiments were conducted in both text-based simulations and a real office environment over the course of a month and a half. Our experiments demonstrate the effectiveness of the proposed framework, showing that AssistantX can reactively respond to user instructions, actively adjust strategies to adapt to contingencies, and proactively seek assistance from humans to ensure successful task completion. More details and videos can be found at https://assistantx-agent.github.io/AssistantX/.

Large language models (LLMs) are increasingly integrated into autonomous systems, giving rise to a new class of software known as Agentware, where LLM-powered agents perform complex, open-ended tasks in domains such as software engineering, customer service, and data analysis. However, their high autonomy and opaque reasoning processes pose significant challenges for traditional software observability methods. To address this, we introduce the concept of cognitive observability—the ability to recover and inspect the implicit reasoning behind agent decisions. We present Watson, a general-purpose framework for observing the reasoning processes of fast-thinking LLM agents without altering their behavior. Watson retroactively infers reasoning traces using prompt attribution techniques. We evaluate Watson in both manual debugging and automated correction scenarios across the MMLU benchmark and the AutoCodeRover and OpenHands agents on the SWE-bench-lite dataset. In both static and dynamic settings, Watson surfaces actionable reasoning insights and supports targeted interventions, demonstrating its practical utility for improving transparency and reliability in Agentware systems.

This work examines the integration of large language models (LLMs) into multi-agent simulations by replacing the hard-coded programs of agents with LLM-driven prompts. The proposed approach is showcased in the context of two examples of complex systems from the field of swarm intelligence: ant colony foraging and bird flocking. Central to this study is a toolchain that integrates LLMs with the NetLogo simulation platform, leveraging its Python extension to enable communication with GPT-4o via the OpenAI API. This toolchain facilitates prompt-driven behavior generation, allowing agents to respond adaptively to environmental data. For both example applications mentioned above, we employ both structured, rule-based prompts and autonomous, knowledge-driven prompts. Our work demonstrates how this toolchain enables LLMs to study self-organizing processes and induce emergent behaviors within multi-agent environments, paving the way for new approaches to exploring intelligent systems and modeling swarm intelligence inspired by natural phenomena. We provide the code, including simulation files and data at https://github.com/crjimene/swarm_gpt.

The Internet of Things (IoT), particularly its industrial subset Industrial IoT (IIoT), presents a critical attack surface due to its interconnected nature. As emerging threats exploit IoT edge networks, there is a growing demand for anomaly detection systems capable of addressing zero-day attacks while providing explainable predictions. Existing machine learning (ML) and deep learning (DL) methods often lack explainability, sensitivity, absence of a large language model (LLM) agent for adaptive detection and struggle with unseen zero-day threats. Motivated by these challenges, we introduce Anomaly-Agent, a novel LLM-powered explainable anomaly detection framework for IoT/IIoT edge environments. Anomaly-Agent leverages a reasoning-followed-by-action pipeline, integrating domain tools, external knowledge retrieval, and memory-augmented decisions to detect and explain anomalies. Unlike static ML/DL models, Anomaly-Agent adapts to evolving zero-day threats and supports sensitivity customization. We evaluated Anomaly-Agent on the Edge-IIoTset (IIoT-specific) and CIC-IoT2023 (general IoT) datasets; it achieved accuracies of 0.96 and 0.89, respectively, with a false alarm rate (FAR) below 0.04. It also attains a recall of 0.65 for zero-day attacks, surpassing traditional ML models and LLM baselines including GPT-4o, Claude 3.5, and GPT-4o-mini. Anomaly-Agent outperformed GPT-4o and Claude 3.5 due to their reliance on generic prompting, which limits performance to 64%–70% multiclass $F1$ -score on CIC-IoT2023. Their high FAR of 10%–13% stems from misclassifying benign edge traffic as malicious. It also surpasses GPT-4o-mini, where token constraints reduce accuracy to 58% for multiclass tasks. The agent’s performance benefits from integration with Shapley additive explanation (SHAP), enhancing transparency and trust. While demonstrating strong performance, Anomaly-Agent faces inherent challenges in latency in complex scenarios and adversarial robustness that guide future improvements. These results demonstrate Anomaly-Agent’s robustness and interpretability, offering a viable path toward resilient, LLM-driven IoT/IIoT security solutions.

This paper presents a large language model (LLM)-based system for autonomous maintenance in manufacturing facilities. While many machine alarms are interpreted with existing manuals, understqanding and acting on these instructions of all facilities remains a challenge for operators. The proposed system processes user inputs including error codes, identifies corresponding procedures from manuals, and decomposes them into structured action sequences. These sequences include action, user interface target, preconditions, and expected outcomes, and are executed by agent capable of interacting with Human–Machine Interfaces (HMIs). The proposed system is built on an LLM-powered multi-agent framework comprising four agents: a chatbot, solution_finder, actor, and supervisor. Each agent operates based on role-specific prompts that define their responsibilities and decision rules. Instead of relying on predefined rule sets, the system interprets unfamiliar or previously unseen alarms by reasoning over machine manuals and context, enabling flexible and scalable maintenance. The system was implemented on a HMI system of CNC machine tools and successfully performed automatic responses to selected alarms. Prompt-based control ensure adaptability to other machines, and the use of a local LLM maintains data security. This approach enables general-purpose, self-directed maintenance with minimal operator intervention.

Although LLM-based agents, powered by Large Language Models (LLMs), can use external tools and memory mechanisms to solve complex real-world tasks, they may also introduce critical security vulnerabilities. However, the existing literature does not comprehensively evaluate attacks and defenses against LLM-based agents. To address this, we introduce Agent Security Bench (ASB), a comprehensive framework designed to formalize, benchmark, and evaluate the attacks and defenses of LLM-based agents, including 10 scenarios (e.g., e-commerce, autonomous driving, finance), 10 agents targeting the scenarios, over 400 tools, 27 different types of attack/defense methods, and 7 evaluation metrics. Based on ASB, we benchmark 10 prompt injection attacks, a memory poisoning attack, a novel Plan-of-Thought backdoor attack, 4 mixed attacks, and 11 corresponding defenses across 13 LLM backbones. Our benchmark results reveal critical vulnerabilities in different stages of agent operation, including system prompt, user prompt handling, tool usage, and memory retrieval, with the highest average attack success rate of 84.30\%, but limited effectiveness shown in current defenses, unveiling important works to be done in terms of agent security for the community. We also introduce a new metric to evaluate the agents' capability to balance utility and security. Our code can be found at https://github.com/agiresearch/ASB.

Agents powered by large language models have shown remarkable abilities in solving complex tasks. However, most agent systems remain reactive, limiting their effectiveness in scenarios requiring foresight and autonomous decision-making. In this paper, we tackle the challenge of developing proactive agents capable of anticipating and initiating tasks without explicit human instructions. We propose a novel data-driven approach for this problem. Firstly, we collect real-world human activities to generate proactive task predictions. These predictions are then labeled by human annotators as either accepted or rejected. The labeled data is used to train a reward model that simulates human judgment and serves as an automatic evaluator of the proactiveness of LLM agents. Building on this, we develop a comprehensive data generation pipeline to create a diverse dataset, ProactiveBench, containing 6,790 events. Finally, we demonstrate that fine-tuning models with the proposed ProactiveBench can significantly elicit the proactiveness of LLM agents. Experimental results show that our fine-tuned model achieves an F1-Score of 66.47% in proactively offering assistance, outperforming all open-source and close-source models. These results highlight the potential of our method in creating more proactive and effective agent systems, paving the way for future advancements in human-agent collaboration.

As autonomous agents powered by large language models (LLMs) continue to demonstrate potential across various assistive tasks, ensuring their safe and reliable behavior is crucial for preventing unintended consequences. In this work, we introduce CIP, a novel technique that leverages causal influence diagrams (CIDs) to identify and mitigate risks arising from agent decision-making. CIDs provide a structured representation of cause-and-effect relationships, enabling agents to anticipate harmful outcomes and make safer decisions. Our approach consists of three key steps: (1) initializing a CID based on task specifications to outline the decision-making process, (2) guiding agent interactions with the environment using the CID, and (3) iteratively refining the CID based on observed behaviors and outcomes. Experimental results demonstrate that our method effectively enhances safety in both code execution and mobile device control tasks.

This paper proposes a novel Neural Architecture Search (NAS) framework powered by Large Language Model (LLM)-based autonomous agents. The method integrates self-consistent architecture generation, role-based agent collaboration, and a memory stream for iterative refinement. By simulating structured dialogues between a Strategic Director and a Technical Analyst, the system enables autonomous design exploration and decision-making without human input. A case study on short-term electricity load forecasting demonstrates the framework’s effectiveness, where the final model—developed through iterative reasoning—achieves a 9.2% reduction in MAE, 7.4% reduction in RMSE, and 32% fewer parameters compared to baseline architectures. The results highlight the framework’s ability to generate efficient, high-performing models while maintaining adaptability and interpretability. This approach offers a scalable solution for automated deep learning design and can be extended to other domains requiring efficient and context-aware model optimisation.

ABSTRACT Powered by the emerging large language models (LLMs), autonomous geographic information system (GIS) agents can perform spatial analyses and cartographic tasks. However, a research gap exists in enabling these agents to autonomously discover and retrieve the necessary data for spatial analysis. This study proposes an autonomous GIS agent framework capable of retrieving required geospatial data by generating, executing, and debugging programs. The framework, with an LLM-driven decision core, selects data sources from a predefined list and fetches data using source-specific handbooks that document metadata and data retrieval details. Designed in a plug-and-play style, the framework allows human users or automated data crawlers to add new sources by creating additional handbooks. A prototype agent based on the framework is developed and released as a QGIS plugin and a Python program. Experiment results demonstrate its capability of retrieving data from various sources, including OpenStreetMap, administrative boundaries and demographic data from the U.S. Census Bureau, satellite basemaps from ESRI World Imagery, global digital elevation model (DEM) from OpenTopography.org, weather data from a commercial provider, and the COVID-19 case data from the NYTimes GitHub. This study is among the first attempts to develop an autonomous GIS agent for geospatial data retrieval.

A GUI agent is an intelligent autonomous system that operates graphical interfaces to complete tasks and deliver services to users. Recent advances in large vision models (LVMs) have boosted GUI agents' flexibility, robustness, and task success rates, but at the cost of high response latency that limits enterprise use. We propose a hybrid LVM GUI agent for Walmart MyAssistant Application: the LVM is invoked only at critical decision points, while routine visual actions are offloaded to lightweight computer vision. Guided by semantics from LLMpredicted action steps, the agent retrieves candidate icons from a curated icon library via semantic search and applies template matching to score tensor similarity against screenshot regions, efficiently estimating click targets. This design cuts end-to-end latency while preserving high success rates, enabling scalable, responsive enterprise application and moving LLM-powered assistants toward broader AGI capabilities.

Lack of transparency and information reliability in supply chain management have been persistent challenges. Based on the LangChain and LangGraph frameworks, this research proposed a Large Language Model (LLM)-based Multi-Agent System (MAS) specifically designed to enhance information reliability and transparency in construction supply chain coordination. A prototype system composed of multiple autonomous agents was designed and developed capable of working collaboratively, sharing information, and supporting decision-making. The system comprises Supplier Agents and General Contractor Agents capable of engaging in natural language interactions. These agents coordinate the supply chain by facilitating communication about material deliveries and project progress. The prototype demonstrated the potential of LLM-based MAS in improving supply chain transparency and reliability. This research not only validated the feasibility of applying large language models in automated supply chain coordination but also offered insights for the design and implementation of future systems.

The cyber-security ecosystem is evolving very fast, with Artificial Intelligence (AI) giving rise to both highly defensive and more sophisticated forms of Advanced Persistent Threats (APTs). AI-powered APTs are a new breed of intelligent, adaptive and self-learning cyber attackers that can autonomously use vulnerabilities, evade detection and continue operating within networks. Organizations in their turn are moving towards the shift between stationary, rule-based control and fully autonomous defensive agents able to conduct continuous monitoring, predict the threat, interrupt the attack real-time, and actively respond. It is this paper that examines the new paradigm of Agent-vs-Agent Cyber Warfare, where autonomous AI defenses indirectly respond to AI-driven APTs on dynamic digital platforms. We describe the architecture of the offensive APT agents based on AI, analyze defensive multi-agent systems (MAS), and suggest a proactive cyber-battlefield model, based on reinforcement learning (RL), large language models (LLM), and self-evolving threat intelligence. Lastly, we outline constraints, ethical aspects, and the way forward with regard to obtaining digital ecosystems in an era of autonomous cyber warfare.

While the recommendation system (RS) has advanced significantly through deep learning, current RS approaches usually train and fine-tune models on task-specific datasets, limiting their generalizability to new recommendation tasks and their ability to leverage external knowledge due to model scale and data size constraints. Thus, we designed an LLM-powered autonomous recommender agent, RecMind, which is capable of leveraging external knowledge, utilizing tools with careful planning to provide zero-shot personalized recommendations. We propose a Self-Inspiring algorithm to improve the planning ability. At each intermediate step, the LLM self-inspires to consider all previously explored states to plan for the next step. This mechanism greatly improves the model's ability to comprehend and utilize historical information in planning for recommendation. We evaluate RecMind's performance in various recommendation scenarios. Our experiment shows that RecMind outperforms existing zero/few-shot LLM-based recommendation baseline methods in various tasks and achieves comparable performance to a fully trained recommendation model P5.

Auditing recommendation systems has attracted growing attention due to increasing concerns over filter bubbles, unfairness, and data misuse. A common approach is sock-puppet auditing, where autonomous agents interact with platforms to reveal risks. However, existing approaches rely on hard-coded agents, lacking adaptability to dynamic GUI layouts and generating behaviors far from those of real users, limiting the comprehensiveness and representativeness of assessment. To address these issues, we introduce AuditAgent, an LLM-powered GUI-agent framework for risk auditing. AuditAgent simulates realistic user preferences and performs adaptive, human-like interactions on recommendation platforms. This design enables more thorough and faithful auditing, providing comprehensive assessments across multiple risk dimensions, including filter bubbles, unfairness, and data misuse.

The successful integration of large language models (LLMs) into laboratory workflows has demonstrated robust capabilities in natural language processing, autonomous task execution, and collaborative problem-solving. This offers an exciting opportunity to realize the dream of autonomous chemical research on demand. Here, we report a robotic AI chemist powered by a hierarchical multiagent system, ChemAgents, based on an on-board Llama-3.1-70B LLM, capable of executing complex, multistep experiments with minimal human intervention. It operates through a Task Manager agent that interacts with human researchers and coordinates four role-specific agents─Literature Reader, Experiment Designer, Computation Performer, and Robot Operator─each leveraging one of four foundational resources: a comprehensive Literature Database, an extensive Protocol Library, a versatile Model Library, and a state-of-the-art Automated Lab. We demonstrate its versatility and efficacy through six experimental tasks of varying complexity, ranging from straightforward synthesis and characterization to more complex exploration and screening of experimental parameters, culminating in the discovery and optimization of functional materials. Additionally, we introduce a seventh task, where ChemAgents is deployed in a new robotic chemistry lab environment to autonomously perform photocatalytic organic reactions, highlighting ChemAgents's scalability and adaptability. Our multiagent-driven robotic AI chemist showcases the potential of on-demand autonomous chemical research to accelerate discovery and democratize access to advanced experimental capabilities across academic disciplines and industries.

Fully autonomous teams of LLM-powered AI agents are emerging that collaborate to perform complex tasks for users. What challenges do developers face when trying to build and debug these AI agent teams? In formative interviews with five AI agent developers, we identify core challenges: difficulty reviewing long agent conversations to localize errors, lack of support in current tools for interactive debugging, and the need for tool support to iterate on agent configuration. Based on these needs, we developed an interactive multi-agent debugging tool, AGDebugger, with a UI for browsing and sending messages, the ability to edit and reset prior agent messages, and an overview visualization for navigating complex message histories. In a two-part user study with 14 participants, we identify common user strategies for steering agents and highlight the importance of interactive message resets for debugging. Our studies deepen understanding of interfaces for debugging increasingly important agentic workflows.

Autonomous agent systems powered by Large Language Models (LLMs) have demonstrated promising capabilities in automating complex tasks. However, current evaluations largely rely on success rates without systematically analyzing the interactions, communication mechanisms, and failure causes within these systems. To bridge this gap, we present a benchmark of 34 representative programmable tasks designed to rigorously assess autonomous agents. Using this benchmark, we evaluate three popular open-source agent frameworks combined with two LLM backbones, observing a task completion rate of approximately 50%. Through in-depth failure analysis, we develop a three-tier taxonomy of failure causes aligned with task phases, highlighting planning errors, task execution issues, and incorrect response generation. Based on these insights, we propose actionable improvements to enhance agent planning and self-diagnosis capabilities. Our failure taxonomy, together with mitigation advice, provides an empirical foundation for developing more robust and effective autonomous agent systems in the future.

Recent advances in large language models (LLMs) and agent-based orchestration are transforming automated knowledge graph (KG) creation as well as robust question answering in complex domains. We present a modular, multi-agent system that extracts, integrates, and reasons over diverse NoSQL movie data, powered by state-of-the-art LLMs such as GPT-4.1. Our architecture converts unstructured plots, cast/crew metadata, and numeric attributes into high-fidelity KGs - enabling both natural language and programmatic queries. To maximize reliability and flexibility, the system unifies multiple retrieval strategies - keyword search, vector similarity, knowledge graph querying, and summarization - each deployed as an autonomous pipeline. Parallel orchestration via LangGraph supports adaptive engine selection, concurrent execution, and robust answer verification with LLM ensemble “jury” scoring. Critically, the framework features comprehensive observability, allowing detailed monitoring and analysis of agent decisions, pipeline performance, and query outcomes. By treating each retrieval method and LLM as a specialized agent, our approach delivers scalable, explainable, and highly accurate results (up to 97%), significantly surpassing monolithic solutions. This agentic, observable architecture paves the way for next-generation autonomous analytics, integration, and decision support across data-rich domains.

The advent of Large Language Models (LLMs) has created new opportunities for the automation of scientific research spanning both experimental processes and computational simulations. This study explores the feasibility of constructing an autonomous simulation agent (ASA) powered by LLMs through prompt engineering and automated program design to automate the entire simulation research process according to a human-provided research plan. This process includes experimental design, remote upload and simulation execution, data analysis, and report compilation. Using a well-studied simulation problem of polymer chain conformations as a test case, we assessed the long-task completion and reliability of ASAs powered by different LLMs, including GPT-4o, Claude-3.5, etc. Our findings revealed that ASA-GPT-4o achieved near-flawless execution on designated research missions, underscoring the potential of methods like ASA to achieve automation in simulation research processes to enhance research efficiency. The outlined automation can be iteratively performed for up to 20 cycles without human intervention, illustrating the potential of ASA for long-task workflow automation. Additionally, we discussed the intrinsic traits of ASA in managing extensive tasks, focusing on self-validation mechanisms, and the balance between local attention and global oversight.

Robots are no longer just tools—they are becoming social agents that can shape how we engage with culture. This study examines what influences visitors’ perceptions of an autonomous museum guide robot, focusing not only on technical capabilities but also on human-centered factors. In a maritime exhibition, 34 participants interacted with a fully autonomous, LLM-powered robot acting as a museum guide. Using self-report questionnaires, we explored how individual differences - age and prior experience with robots — interacted with experimental conditions to shape participants’ impressions of the robot. Our findings suggest that these personal factors significantly affect how visitors evaluate the robot, suggesting that effective design must reflect the diversity of users’ experiences and expectations. By acknowledging the complexity of human-robot interaction, we move closer to creating robotic guides that are not only functional but also socially attuned.

Recently, there has been an emergence of employing LLM-powered agents as believable human proxies, based on their remarkable decision-making capability. However, existing studies mainly focus on simulating human dialogue. Human non-verbal behaviors, such as item clicking in recommender systems, although implicitly exhibiting user preferences and could enhance the modeling of users, have not been deeply explored. The main reasons lie in the gap between language modeling and behavior modeling, as well as the incomprehension of LLMs about user-item relations. To address this issue, we propose AgentCF for simulating user-item interactions in recommender systems through agent-based collaborative filtering. We creatively consider not only users but also items as agents, and develop a collaborative learning approach that optimizes both kinds of agents together. Specifically, at each time step, we first prompt the user and item agents to interact autonomously. Then, based on the disparities between the agents' decisions and real-world interaction records, user and item agents are prompted to reflect on and adjust the misleading simulations collaboratively, thereby modeling their two-sided relations. The optimized agents can also propagate their preferences to other agents in subsequent interactions, implicitly capturing the collaborative filtering idea. Overall, the optimized agents exhibit diverse interaction behaviors within our framework, including user-item, user-user, item-item, and collective interactions. The results show that these agents can demonstrate personalized behaviors akin to those of real-world individuals, sparking the development of next-generation user behavior simulation.

No abstract available

Large Language Models (LLM) are increasingly being explored for problem-solving tasks. However, their strategic planning capability is often viewed with skepticism. Recent studies have incorporated the Monte Carlo Tree Search (MCTS) algorithm to augment the planning capacity of LLM. Despite its potential, MCTS relies on extensive sampling simulations to approximate the true reward distribution, which leads to two primary issues. Firstly, MCTS is effective for tasks like the Game of Go, where simulation results can yield objective rewards (e.g., 1 for a win and 0 for a loss). However, for tasks such as question answering, the result of a simulation is the answer to the question, which cannot yield an objective reward without the ground truth. Secondly, obtaining statistically significant reward estimations typically requires a sample size exceeding 30 simulations, resulting in excessive token usage and time consumption. To address these challenges, we present the Multi-Agent System with Tactical Execution and Reasoning using LLM Specialized MCTS (MASTER), a novel framework that coordinates agent recruitment and communication through LLM specialized MCTS. This system autonomously adjusts the number of agents based on task complexity and ensures focused communication among them. Comprehensive experiments across various tasks demonstrate the effectiveness of our proposed framework. It achieves 76% accuracy on HotpotQA and 80% on WebShop, setting new state-of-the-art performance on these datasets.

No abstract available

Aiming at the problem that the increasing number of automated guided vehicles (AGVs) will lead to more frequent conflicts between AGVs. In this paper, a conflict-free path planning model for multi-AGV is established, aiming to minimize the blocking rate of AGVs between the quay crane and the yard crane, considering the travel speed, operation time, and conflict distance of AGVs. An architecture of AGV’s system based on Multi-Agent System (MAS) is designed, the improved interactive protocol based on blackboard model is used as the communication method of AGV, the improved acceleration control method is combined with the AGV priority determination method based on time cost as the negotiation strategy of AGV, the improved Dijkstra algorithm calculates the conflict-free path of each AGV. By comparing the acceleration control method based on MAS with the speed control method based on MAS and the task priority control method, the effectiveness of this method for solving multiple AGVs conflict-free path planning in automated terminals is verified.

No abstract available

No abstract available

No abstract available

No abstract available

Emergency department (ED) overcrowding and the complexity of rapid decision-making in critical care settings pose significant challenges to healthcare systems worldwide. While clinical decision support systems (CDSS) have shown promise, the integration of large language models (LLMs) offers new possibilities for enhancing triage accuracy and clinical decisionmaking. This study presents an LLM-driven CDSS designed to assist ED physicians and nurses in patient triage, treatment planning, and overall emergency care management. We developed a multi-agent CDSS utilizing Llama-3-70b as the base LLM, orchestrated by CrewAI and Langchain. The system comprises four AI agents emulating key ED roles: Triage Nurse, Emergency Physician, Pharmacist, and ED Coordinator. It incorporates the Korean Triage and Acuity Scale (KTAS) for triage assessment and integrates with the RxNorm API for medication management. The model was evaluated using the Asclepius dataset, with performance assessed by a clinical emergency medicine specialist. The CDSS demonstrated high accuracy in triage decisionmaking compared to the baseline of a single-agent system. Furthermore, the system exhibited strong performance in critical areas, including primary diagnosis, critical findings identification, disposition decision-making, treatment planning, and resource allocation. Our multi-agent CDSS demonstrates significant potential for supporting comprehensive emergency care management. By leveraging state-of-the-art AI technologies, this system offers a scalable and adaptable tool that could enhance emergency medical care delivery, potentially alleviating ED overcrowding and improving patient outcomes. This work contributes to the growing field of AI applications in emergency medicine and offers a promising direction for future research and clinical implementation.

In this paper, we introduce a fault-tolerant multi-agent reinforcement learning framework called SERT-DQN to optimize the operations of UAVs with UGV central control in coverage path planning missions. Our approach leverages dual learning systems that combine individual agent autonomy with centralized strategic planning, thus enhancing the efficiency of cooperative path planning missions. This framework is designed for high performance in environments with fault uncertainty detected and operational challenges such as interruptions in connectivity and compromised sensor reliability. With the integration of an innovative communication system between agents, our system appropriately handles both static and dynamic environments. Also, we introduce similarity-based shared experience replay to attain faster convergence and sample efficiency in the multi-agent system. The architecture is specially designed to respond adaptively to such irregularities by effectively showing enhanced resilience in scenarios where data integrity is impaired due to faults or the UAV faces disruptions. Simulation results indicate that our fault tolerance algorithms are very resilient and do indeed improve mission outcomes, especially under dynamic and highly uncertain operating conditions. This approach becomes critical for the most recent sensor-based research in autonomous systems.

No abstract available

Manufacturing systems are undergoing systematic change facing the trade-off between the customer's needs and the economic and ecological pressure. Especially assembly systems must be more flexible due to many product generations or unpredictable material and demand fluctuations. As a solution line-less mobile assembly systems implement flexible job routes through movable multi-purpose resources and flexible transportation systems. Moreover, a completely reactive rearrangeable layout with mobile resources enables reconfigurations without interrupting production. A scheduling that can handle the complexity of dynamic events is necessary to plan job routes and control transportation in such an assembly system. Conventional approaches for this control task require exponentially rising computational capacities with increasing problem sizes. Therefore, the contribution of this work is an algorithm to dynamically solve the integrated problem of layout optimization and scheduling in line-less mobile assembly systems. The proposed multi agent deep reinforcement learning algorithm uses proximal policy optimization and consists of a decoder and encoder, allowing for various-sized system state descriptions. A simulation study shows that the proposed algorithm performs better in 78% of the scenarios compared to a random agent regarding the makespan optimization objective. This allows for adaptive optimization of line-less mobile assembly systems that can face global challenges.

No abstract available

Abstract A distributed multi-project scheduling problem is considered, in which several projects share scarce resources, and a planning department (planner) is responsible for allocating the resources among the projects. Information asymmetry and heterogeneous resources are assumed to be due to the geographical distribution of the planner and the projects. The projects compete for the limited global resources to maximize their local benefit, such that they may lie or overstate resource importance to the planner. In this paper, a multi-agent system is developed to address this problem due to the concerns of private information and highly autonomous nature of project agents, which makes a central coordination approach unsuitable. Different from previous work, a project agent may employ the lying strategy to increase its possibility of winning the desired resource, while the planner can adopt an integrity policy to penalize this behaviour. Another main contribution is that a heuristic procedure is designed and combined with an argumentation-based approach for this multi-agent system that can improve computation efficiency. Finally, the proposed combined multi-agent system is compared with a central coordination algorithm to demonstrate its efficacy. Numerical experiments show that the combined multi-agent system is more effective in exploration. It outperforms the central coordination algorithm for problems of a larger scale, especially those with a tighter global resource constraint. Experimental results also reveal that the proper integrity policy could considerably reduce the negative effect of dishonesty of the project agents on the global objective by eliminating the potential to benefit from lying.

No abstract available

UAV-based air-ground integrated computing networks (AGIN) have gained significant traction in remote areas for the Power Internet of Things (PIoT). This paper considers an AGIN-PIoT, where computing tasks generated by ground PIoT devices are offloaded to aerial UAVs that perform edge computing. Jointly optimizing task offloading and UAV trajectory poses challenges such as many decision variables, information uncertainty, and long-term queue delay constraints. Due to the limited battery capacity of PIoT devices and UAVs, our objective is to minimize system energy consumption under long-term queue delay constraints by jointly optimizing task offloading, trajectory planning, and computing resource assignment. In light of Lyapunov optimization, we decompose the original challenging optimization problem into two sub-problems: (1) task offloading and UAV trajectory planning and (2) aerial edge resource allocation. Accordingly, we develop a multi-agent deep reinforcement learning-based algorithm called AGIN-MADDPG for the former to achieve the maximum accumulative reward and propose a greedy solution for the latter. Extensive experiments and numerical results demonstrate that our approach can avoid the problem of gradient vanishing and outperforms other benchmark methods in terms of power consumption, task backlog, queue delay, and system throughput.

Multi-agent Pathfinding (MAPF) problem generally asks to find a set of conflict-free paths for a set of agents confined to a graph and is typically solved in a centralized fashion. Conversely, in this work, we investigate the decentralized MAPF setting, when the central controller that possesses all the information on the agents' locations and goals is absent and the agents have to sequentially decide the actions on their own without having access to the full state of the environment. We focus on the practically important lifelong variant of MAPF, which involves continuously assigning new goals to the agents upon arrival to the previous ones. To address this complex problem, we propose a method that integrates two complementary approaches: planning with heuristic search and reinforcement learning through policy optimization. Planning is utilized to construct and re-plan individual paths. We enhance our planning algorithm with a dedicated technique tailored to avoid congestion and increase the throughput of the system. We employ reinforcement learning to discover the collision avoidance policies that effectively guide the agents along the paths. The policy is implemented as a neural network and is effectively trained without any reward-shaping or external guidance. We evaluate our method on a wide range of setups comparing it to the state-of-the-art solvers. The results show that our method consistently outperforms the learnable competitors, showing higher throughput and better ability to generalize to the maps that were unseen at the training stage. Moreover our solver outperforms a rule-based one in terms of throughput and is an order of magnitude faster than a state-of-the-art search-based solver. The code is available at https://github.com/AIRI-Institute/learn-to-follow.

Task scheduling (TS) and multi-agent-path-finding (MAPF) are two cruxes of pickup-and-delivery in automated warehouses. In this paper, the two cruxes are optimized simultaneously. Firstly, the system model, task model, and path model are established, respectively. Then, a task scheduling algorithm based on enhanced HEFT, a heuristic MAPF algorithm and a TS- MAPF algorithm are proposed to solve this combinatorial optimization problem. In EHEFT, a novel rank priority rule is used to determine task sequencing and task allocation. In MAPF algorithm, a CBS algorithm with priority rules is designed for path search. Subsequently, the TS-MAPF algorithm which combines EHEFT and MAPF is proposed. Finally, the proposed algorithms are tested separately against relevant typical algorithms at different scales. The experimental results indicate that the proposed algorithms exhibited good performance.

Efficient coordination and planning is essential for large-scale multi-agent systems that collaborate in a shared dynamic environment. Heuristic search methods or learning-based approaches often lack the guarantee on correctness and performance. Moreover, when the collaborative tasks contain both spatial and temporal requirements, e.g., as linear temporal logic (LTL) formulas, formal methods provide a verifiable framework for task planning. However, since the planning complexity grows exponentially with the number of agents and the length of the task formula, existing studies are mostly limited to small artificial cases. To address this issue, a new planning paradigm is proposed in this work for system-wide temporal task formulas that are released online and continually. It avoids two common bottlenecks in the traditional methods, i.e., (a) the direct translation of the complete task formula to the associated Büchi automaton and (b) the synchronized product between the Büchi automaton and the transition models of all agents. Instead, an adaptive planning algorithm is proposed, which computes the product of relaxed partially ordered sets (R-posets) on-the-fly and assigns these subtasks to the agents subject to the ordering constraints. It is shown that the first valid plan can be derived with a polynomial time and memory complexity with respect to the system size and the formula length. Our method can take into account task formulas with a length of more than 400 and a fleet with more than 400 agents, while most existing methods fail at the formula length of 25 within a reasonable duration. The proposed method is validated on large fleets of service robots in both simulation and hardware experiments.

This study investigates the implementation of LLM agents in smart city management, leveraging both the inherent language processing abilities of LLMs and the distributed problem solving capabilities of multi-agent systems for the improvement of urban decision making processes. A multi-agent system architecture combines LLMs with existing urban information systems to process complex queries and generate contextually relevant responses for urban planning and management. The research is focused on three main hypotheses testing: (1) LLM agents’ capability for effective routing and processing diverse urban queries, (2) the effectiveness of Retrieval-Augmented Generation (RAG) technology in improving response accuracy when working with local knowledge and regulations, and (3) the impact of integrating LLM agents with existing urban information systems. Our experimental results, based on a comprehensive validation dataset of 150 question–answer pairs, demonstrate significant improvements in decision support capabilities. The multi-agent system achieved pipeline selection accuracy of 94–99% across different models, while the integration of RAG technology improved response accuracy by 17% for strategic development queries and 55% for service accessibility questions. The combined use of document databases and service APIs resulted in the highest performance metrics (G-Eval scores of 0.68–0.74) compared to standalone LLM responses (0.30–0.38). Using St. Petersburg’s Digital Urban Platform as a testbed, we demonstrate the practical applicability of this approach to create integrated city management systems with support complex urban decision making processes. This research contributes to the growing field of AI-enhanced urban management by providing empirical evidence of LLM agents’ effectiveness in processing heterogeneous urban data and supporting strategic planning decisions. Our findings suggest that LLM-based multi-agent systems can significantly enhance the efficiency and accuracy of urban decision making while maintaining high relevance in responses.

Fast moving unmanned aerial vehicles (UAVs) are well suited for aerial surveillance, but are limited by their battery capacity. To increase their endurance, UAVs can be refueled on slow moving unmanned ground vehicles (UGVs). This cooperative routing of UAV-UGV multi-agent system to survey vast regions within their speed and fuel constraints is a computationally challenging problem, but can be simplified with heuristics. In this study, we utilize heuristic approaches to obtain feasible and near-optimal solutions to the problem, leveraging the fuel limitations of the UAV with the minimum set cover algorithm to identify the UGV refueling points. These refueling stops enable the allocation of mission points to the UAV and UGV. A standard traveling salesman formulation and a vehicle routing formulation with time windows, dropped visits, and capacity constraints are used to solve for the UGV and UAV route, respectively. Experimental validation on a small-scale testbed (http://tiny.cc/vancvz) underscores the effectiveness of our multi-agent approach.

Advancements in Intelligent Transportation Systems (ITS) have led to innovative solutions for planning optimization, efficiency enhancement, and resource allocation in transportation networks, which are demonstrated in applications such as smart parking lot management and electric vehicle (EV) charging station allocation, where improved decision-making and system-wide optimization have been achieved. However, as these systems evolve, the demand for better adaptability and coordination continues to grow to maximize their overall effectiveness and efficiency. To achieve this, we propose the Multi-Personality Multi-Agent Meta-Reinforcement Learning (MPMA-MRL) framework. This approach incorporates multiple meta-trained, meta-tested explainable personality policies, which are deployed to each agent. A personality selector is trained and deployed on each agent to optimize the overall performance. MPMA-MRL is superior than traditional methods in terms of the adaptability and coordination in ITS by leveraging improved information from the environment, more practical coordination among agents, faster adaptation speed to intermediate tasks, and more appropriate allocation and planning. The proposed framework is evaluated in the applications of parking lot optimization and EV charging station allocation. Its broader impact on multi-agent smart systems is analyzed to demonstrate its generalizability. The results demonstrate that in parking lot optimization, MPMA-MRL significantly reduces the time required to direct all vehicles to available parking spots. In EV charging station allocation, MPMA-MRL effectively minimizes waiting times at charging stations. Moreover, in both applications, MPMA-MRL exhibits enhanced adaptability to previously unseen scenarios, improving its applicability.

We tackle the challenging problem of multi-agent cooperative motion planning for complex tasks described using signal temporal logic (STL), where robots can have nonlinear and nonholonomic dynamics. Existing methods in multi-agent motion planning, especially those based on discrete abstractions and model predictive control (MPC), suffer from limited scalability with respect to the complexity of the task, the size of the workspace, and the planning horizon. We present a method based on {\em timed waypoints\/} to address this issue. We show that timed waypoints can help abstract nonlinear behaviors of the system as safety envelopes around the reference path defined by those waypoints. Then the search for waypoints satisfying the STL specifications can be inductively encoded as a mixed-integer linear program. The agents following the synthesized timed waypoints have their tasks automatically allocated, and are guaranteed to satisfy the STL specifications while avoiding collisions. We evaluate the algorithm both in simulation and on the Robotarium platform. Results show that it supports multi-agent planning from complex specification over long planning horizons, and significantly outperforms state-of-the-art abstraction-based and MPC-based motion planning tools.

No abstract available

No abstract available

No abstract available

We study the problem of safe multi-agent motion planning in cluttered environments. Existing multi-agent reinforcement learning-based motion planners only provide approximate safety enforcement. We propose a safe reinforcement learning algorithm that leverages single-agent reinforcement learning for target regulation and a subsequent convex optimization-based filtering that ensures the collective safety of the system. Our approach yields a safe, real-time implementable multi-agent motion planner that is simpler to train and enforces safety as hard constraints. Our approach can handle state and control constraints on the agents, and enforce collision avoidance among themselves and with static obstacles in the environment. Numerical simulations and hardware experiments show the efficacy of the approach.

One of the major research topics in unmanned aerial vehicle (UAV) collaborative control systems is the problem of multi-UAV target assignment and path planning (MUTAPP). It is a complicated optimization problem in which target assignment and path planning are solved separately. However, recalculation of the optimal results is too slow for real-time operations in dynamic environments because of the large number of calculations required. In this paper, we propose an artificial intelligence method named simultaneous target assignment and path planning (STAPP) based on a multi-agent deep deterministic policy gradient (MADDPG) algorithm, which is a type of multi-agent reinforcement learning algorithm. In STAPP, the MUTAPP problem is first constructed as a multi-agent system. Then, the MADDPG framework is used to train the system to solve target assignment and path planning simultaneously according to a corresponding reward structure. The proposed system can deal with dynamic environments effectively as its execution only requires the locations of the UAVs, targets, and threat areas. Real-time performance can be guaranteed as the neural network used in the system is simple. In addition, we develop a technique to improve the training effect and use experiments to demonstrate the effectiveness of our method.

Benefiting from the powerful language expression and planning capabilities of Large Language Models (LLMs), LLM-based autonomous agents have achieved promising performance in various downstream tasks. Recently, based on the development of single-agent systems, researchers propose to construct LLM-based multi-agent systems to tackle more complicated tasks. In this paper, we propose a novel framework, named COPPER , to enhance the collaborative capabilities of LLM-based agents with the self-reflection mechanism. To improve the quality of reflections, we propose to fine-tune a shared reflector, which automatically tunes the prompts of actor models using our counterfactual PPO mechanism. On the one hand, we propose counterfactual rewards to assess the contribution of a single agent’s reflection within the system, alleviating the credit assignment problem. On the other hand, we propose to train a shared reflector, which enables the reflector to generate personalized reflections according to agent roles, while reducing the computational resource requirements and improving training stability. We conduct experiments on three datasets to evaluate the performance of our model in multi-hop question answering, mathematics, and chess scenarios. Experimental results show that COPPER possesses stronger reflection capabilities and exhibits excellent generalization performance across different actor models.

The emergence of Large Language Models (LLMs) like ChatGPT has inspired the development of LLM-based agents capable of addressing complex, real-world tasks. However, these agents often struggle during task execution due to methodological constraints, such as error propagation and limited adaptability. To address this issue, we propose a multi-agent framework based on dynamic Task Decomposition and Agent Generation (TDAG). This framework dynamically decomposes complex tasks into smaller subtasks and assigns each to a specifically generated subagent, thereby enhancing adaptability in diverse and unpredictable real-world tasks. Simultaneously, existing benchmarks often lack the granularity needed to evaluate incremental progress in complex, multi-step tasks. In response, we introduce ItineraryBench in the context of travel planning, featuring interconnected, progressively complex tasks with a fine-grained evaluation system. ItineraryBench is designed to assess agents' abilities in memory, planning, and tool usage across tasks of varying complexity. Our experimental results reveal that TDAG significantly outperforms established baselines, showcasing its superior adaptability and context awareness in complex task scenarios.

No abstract available

Autonomous driving system aims for safe and social-consistent driving through the behavioral integration among interactive agents. However, challenges remain due to multi-agent scene uncertainty and heterogeneous interaction. Current dense and sparse behavioral representations struggle with inefficiency and inconsistency in multi-agent modeling, leading to instability of collective behavioral patterns when integrating prediction and planning (IPP). To address this, we initiate a topological formation that serves as a compliant behavioral foreground to guide downstream trajectory generations. Specifically, we introduce Behavioral Topology (BeTop), a pivotal topological formulation that explicitly represents the consensual behavioral pattern among multi-agent future. BeTop is derived from braid theory to distill compliant interactive topology from multi-agent future trajectories. A synergistic learning framework (BeTopNet) supervised by BeTop facilitates the consistency of behavior prediction and planning within the predicted topology priors. Through imitative contingency learning, BeTop also effectively manages behavioral uncertainty for prediction and planning. Extensive verification on large-scale real-world datasets, including nuPlan and WOMD, demonstrates that BeTop achieves state-of-the-art performance in both prediction and planning tasks. Further validations on the proposed interactive scenario benchmark showcase planning compliance in interactive cases.

Humans collaborate in dynamic and flexible ways. Collaboration requires agents to coordinate their behavior on the fly, sometimes jointly solving a single task together and other times dividing it up into sub-tasks to work on in parallel. We develop Bayesian Delegation, a learning mechanism for decentralized multi-agent coordination that enables agents to rapidly infer the sub-tasks that other agents are working on by inverse planning. These inferences enable agents to determine, in the absence of communication, whether to plan jointly with others or work on complementary sub-tasks. We test this model in a suite of decentralized multi-agent environments inspired by cooking problems. To succeed, agents must coordinate both their high-level plans (sub-task) and their low-level actions (avoiding collisions). Including joint sub-tasks in the prior of Bayesian delegation enables agents to carry out sub-tasks that neither agent can finish independently. The full system outperforms lesioned systems that are missing one or more of these capabilities.

Large Language Models (LLMs) excel in traditional natural language processing tasks but struggle with problems that require complex domain-specific calculations or simulations. While equipping LLMs with external tools to build LLM-based agents can enhance their capabilities, existing approaches lack the flexibility to address diverse and ever-evolving user queries in open domains. Currently, there is also no existing dataset that evaluates LLMs on open-domain knowledge that requires tools to solve. To this end, we introduce OpenAct benchmark to evaluate the open-domain task-solving capability, which is built on human expert consultation and repositories in GitHub. It comprises 339 questions spanning 7 diverse domains that need to be solved with domain-specific methods. In our experiments, even state-of-the-art LLMs and LLM-based agents demonstrate unsatisfactory success rates, underscoring the need for a novel approach. Furthermore, we present OpenAgent, a novel LLM-based agent system that can tackle evolving queries in open domains through autonomously integrating specialized tools from GitHub. OpenAgent employs 1) a hierarchical framework where specialized agents handle specific tasks and can assign tasks to inferior agents, 2) a bi-level experience learning mechanism to learn from both humans' and its own experiences to tackle tool flaws. Experiments demonstrate its superior effectiveness and efficiency, which significantly outperforms baselines. Our data and code are open-source at https://github.com/OpenBMB/OpenAct.

Scientific research increasingly relies on specialized computational tools, yet effectively utilizing these tools requires substantial domain expertise. While large language models show promise in tool automation, they struggle to seamlessly integrate and orchestrate multiple tools for complex scientific workflows. Here we present SciToolAgent, a large language model-powered agent that automates hundreds of scientific tools across biology, chemistry and materials science. At its core, SciToolAgent leverages a scientific tool knowledge graph that enables intelligent tool selection and execution through graph-based retrieval-augmented generation. The agent also incorporates a comprehensive safety-checking module to ensure responsible and ethical tool usage. Extensive evaluations on a curated benchmark demonstrate that SciToolAgent outperforms existing approaches. Case studies in protein engineering, chemical reactivity prediction, chemical synthesis and metal–organic framework screening further demonstrate SciToolAgent’s capability to automate complex scientific workflows, making advanced research tools accessible to both experts and nonexperts. This study presents SciToolAgent, a large language model-based agent that orchestrates scientific tools via a knowledge graph, enabling automated and effective execution of scientific research workflows.

Abstract Motivation The growing complexity of clinical cancer research has fueled a surge in demand for automated bioinformatics tools capable of integrating clinical and genomic data to accelerate discovery efforts. Results We present the Artificial Intelligence Agent for High-Optimization and Precision Medicine (AI-HOPE), an AI-driven system that enables domain experts to conduct integrative data analyses through natural language interactions. Powered by Large Language Models, AI-HOPE interprets user instructions, converts them into executable code, and autonomously analyzes locally stored data. It supports flexible association studies, subset comparisons, clinical prevalence assessments and survival analyses. In addition, AI-HOPE enables global variable scans to identify features significantly associated with a user-defined outcome, making a powerful and intuitive tool for advancing precision medicine research. Importantly, its closed-system design prevents clinical data leakage. To demonstrate its utility, AI-HOPE was applied to The Cancer Genome Atlas data to address two clinical questions. First, it identified significant enrichment of TP53 mutations in late-stage colorectal cancer compared to early-stage cases. Second, it uncovered a strong association between KRAS mutations and poorer progression-free survival in FOLFOX-treated patients. These findings align with established literature and demonstrate AI-HOPE's ability to generate meaningful insights independently, without prior assumptions. By removing programming barriers and simplifying complex analyses, AI-HOPE bridges the gap between data complexity and research needs. With its scalable and adaptable framework, AI-HOPE has the potential to support diverse biomedical research fields, driving innovation and efficiency in translational studies. Availability and implementation The AI-HOPE software and demonstration data is available at https://github.com/Velazquez-Villarreal-Lab/AI-HOPE.

The Model Context Protocol (MCP) is rapidly emerging as a pivotal open standard, designed to enhance agent-tool integration and interoperability, and is positioned to unlock a new era of powerful, interconnected, and genuinely utilitarian agentic AI. However, despite MCP's growing adoption, existing benchmarks often fail to capture real-world agent performance within this new paradigm, leading to a distorted perception of their true operational value and an inability to reliably differentiate proficiencies. To bridge this critical evaluation gap, we introduce MCP-AgentBench—a comprehensive benchmark specifically engineered to rigorously assess language agent capabilities in MCP-mediated tool interactions. Core contributions of MCP-AgentBench include: the establishment of a robust MCP testbed comprising 33 operational servers with 188 distinct tools; the development of a benchmark featuring 600 systematically designed queries distributed across 6 distinct categories of varying interaction complexity; and the introduction of MCP-Eval, a novel outcome-oriented evaluation methodology prioritizing real-world task success. Through extensive empirical evaluation of leading language agents, we provide foundational insights. MCP-AgentBench aims to equip the research community with a standardized and reliable framework to build, validate, and advance agents capable of fully leveraging MCP's transformative benefits, thereby accelerating progress toward truly capable and interoperable AI systems.

ABSTRACT Recent advancements in generative artificial intelligence (AI), particularly Large Language Models (LLMs), offer promising capabilities for spatial analysis. However, their integration with established GIS platforms remains underexplored. In this study, we propose a framework that embeds LLMs into existing GIS platforms, using QGIS as a case study. Our approach leverages LLMs’ reasoning and coding abilities to autonomously generate spatial analysis workflows through an informed agent equipped with comprehensive documentation of key GIS tools and parameters. External tools such as GeoPandas are also incorporated to enhance the system's geoprocessing capabilities. Based on this framework, we developed a ‘GIS Copilot’ that enables users to interact with QGIS using natural language. We evaluated the copilot across over 100 tasks of varying complexity including basic (single tool/layer), intermediate (multistep with guidance), and advanced (multistep without guidance). Results show high success rates for basic and intermediate tasks, with challenges remaining in fully autonomous execution of advanced tasks. The GIS Copilot advances the vision of autonomous GIS by enabling non-experts to perform geospatial analysis with minimal prior knowledge. While full autonomy is not yet achieved, the copilot demonstrates significant potential for simplifying GIS workflows and enhancing decision-making processes.

The integration of experimental technologies with large language models (LLMs) is transforming scientific research. It positions AI as a versatile research assistant rather than a mere problem-solving tool. In the field of power systems, however, managing simulations — one of the essential experimental technologies — remains a challenge for LLMs due to their limited domain-specific knowledge, restricted reasoning capabilities, and imprecise handling of simulation parameters. To address these limitations, this paper proposes a feedback-driven, multi-agent framework. It incorporates three proposed modules: an enhanced retrieval-augmented generation (RAG) module, an improved reasoning module, and a dynamic environmental acting module with an error-feedback mechanism. Validated on 69 diverse tasks from Daline and MATPOWER, this framework achieves success rates of 93.13% and 96.85%, respectively. It significantly outperforms ChatGPT 4o, o1-preview, and the fine-tuned GPT4o, which all achieved a success rate lower than 30% on complex tasks. Additionally, the proposed framework also supports rapid, cost-effective task execution, completing each simulation in approximately 30 seconds at an average cost of 0.014 USD for tokens. Overall, this adaptable framework lays a foundation for developing intelligent LLM-based assistants for human researchers, facilitating power system research and beyond.

No abstract available

Large language models (LLMs) have recently demonstrated remarkable capabilities to comprehend human intentions, engage in reasoning, and design planning-like behavior. To further unleash the power of LLMs to accomplish complex tasks, there is a growing trend to build agent framework that equips LLMs, such as ChatGPT, with tool-use abilities to connect with massive external APIs. In this work, we introduce ModelScope-Agent, a general and customizable agent framework for real-world applications, based on open-source LLMs as controllers. It provides a user-friendly system library, with customizable engine design to support model training on multiple open-source LLMs, while also enabling seamless integration with both model APIs and common APIs in a unified way. To equip the LLMs with tool-use abilities, a comprehensive framework has been proposed spanning over tool-use data collection, tool retrieval, tool registration, memory control, customized model training, and evaluation for practical real-world applications. Finally, we showcase ModelScopeGPT, a real-world intelligent assistant of ModelScope Community based on the ModelScope-Agent framework, which is able to connect open-source LLMs with more than 1000 public AI models and localized community knowledge in ModelScope. The ModelScope-Agent library\footnote{https://github.com/modelscope/modelscope-agent} and online demo\footnote{https://modelscope.cn/studios/damo/ModelScopeGPT/summary} are now publicly available.

No abstract available

No abstract available

Multi-Agent Systems have been widely used for traffic simulation. The modeling of individuals allows indeed to introduce a behavioral diversity which is crucial to obtain realistic simulation outcomes. The recent growth of open geographical databases and related flow information provides an opportunity for enhancing traffic simulators with data automatically retrieved from the real world and updated regularly. We present here TrafficGen, a highly modular platform based on the integration of such open data within a library of rule-based behaviors, in order to provide a versatile decision support tool in traffic.

The continuous change in the manufacturing world is demanding more flexible, responsive and accurate planning tools, which are able to assist the decision-makers to take tactical and strategic decisions on short notice with a high level of confidence. For this purpose, these tools should dynamically explore different operative scenarios in the planning procedure and produce information about key performance indicators. This paper describes the development of an agent-based strategic planner, combining the flexibility of multi-agent systems principles with the optimization capability of a Mixed Integral Programming technique. The tool is integrated in an ecosystem of heterogeneous decision-making systems through an Enterprise Service Bus that also provides access to legacy data.

No abstract available

Automated Planning and Scheduling is among the growing areas in Artificial Intelligence (AI) where mention of LLMs has gained popularity. Based on a comprehensive review of 126 papers, this paper investigates eight categories based on the unique applications of LLMs in addressing various aspects of planning problems: language translation, plan generation, model construction, multi-agent planning, interactive planning, heuristics optimization, tool integration, and brain-inspired planning. For each category, we articulate the issues considered and existing gaps. A critical insight resulting from our review is that the true potential of LLMs unfolds when they are integrated with traditional symbolic planners, pointing towards a promising neuro-symbolic approach. This approach effectively combines the generative aspects of LLMs with the precision of classical planning methods. By synthesizing insights from existing literature, we underline the potential of this integration to address complex planning challenges. Our goal is to encourage the ICAPS community to recognize the complementary strengths of LLMs and symbolic planners, advocating for a direction in automated planning that leverages these synergistic capabilities to develop more advanced and intelligent planning systems. We aim to keep the categorization of papers updated on https://ai4society.github.io/LLM-Planning-Viz/, a collaborative resource that allows researchers to contribute and add new literature to the categorization.

No abstract available

No abstract available

No abstract available

No abstract available

No abstract available

The ongoing fourth Industrial Revolution depends mainly on robust Industrial Cyber-Physical Systems (ICPS). ICPS includes computing (software and hardware) abilities to control complex physical processes in distributed industrial environments. Industrial agents, originating from the well-established multi-agent systems field, provide complex and cooperative control mechanisms at the software level, allowing us to develop larger and more feature-rich ICPS. The IEEE P2660.1 standardisation project, "Recommended Practices on Industrial Agents: Integration of Software Agents and Low Level Automation Functions" focuses on identifying Industrial Agent practices that can benefit ICPS systems of the future. A key problem within this project is identifying the best-fit industrial agent practices for a given ICPS. This paper reports on the design and development of a tool to address this challenge. This tool, called IASelect, is built using graph databases and provides the ability to flexibly and visually query a growing repository of industrial agent practices relevant to ICPS. IASelect includes a front-end that allows industry practitioners to interactively identify best-fit practices without having to write manual queries.

No abstract available

Simulation of smart grid technologies requires a fundamentally new approach to integrated modeling of power systems, energy markets, building technologies, and the plethora of other resources and assets that are becoming part of modern electricity production, delivery, and consumption systems. As a result, the US Department of Energy’s Office of Electricity commissioned the development of a new type of power system simulation tool called GridLAB-D that uses an agent-based approach to simulating smart grids. This paper presents the numerical methods and approach to time-series simulation used by GridLAB-D and reviews applications in power system studies, market design, building control system design, and integration of wind power in a smart grid.

The rise of generative artificial intelligence (GAI), especially with multimodal large language models like GPT‐4o, sparked transformative potential and challenges for learning and teaching. With potential as a cognitive offloading tool, GAI can enable learners to focus on higher‐order thinking and creativity. Yet, this also raises questions about integration into traditional education due to the limited research on learners' interactions with GAI. Some studies with GAI focus on text‐based human–AI interactions, while research on embodied GAI in immersive environments like mixed reality (MR) remains unexplored. To address this, this study investigates interaction dynamics between learners and embodied GAI agents in MR, examining cognitive and socio‐emotional interactions during collaborative learning. We investigated the paired interactive patterns between a student and an embodied GAI agent in MR, based on data from 26 higher education students with 1317 recorded activities. Data were analysed using a multi‐layered learning analytics approach, including quantitative content analysis, sequence analysis via hierarchical clustering and pattern analysis through ordered network analysis (ONA). Our findings identified two interaction patterns: type (1) AI‐led Supported Exploratory Questioning (AISQ) and type (2) Learner‐Initiated Inquiry (LII) group. Despite their distinction in characteristic, both types demonstrated comparable levels of socio‐emotional engagement and exhibited meaningful cognitive engagement, surpassing the superficial content reproduction that can be observed in interactions with GPT models. This study contributes to the human–AI collaboration and learning studies, extending understanding to learning in MR environments and highlighting implications for designing AI‐based educational tools. What is already known about this topic Socio‐emotional interactions are fundamental to cognitive processes and play a critical role in collaborative learning. Generative artificial intelligence (GAI) holds transformative potential for education but raises questions about how learners interact with such technology. Most existing research focuses on text‐based interactions with GAI; there is limited empirical evidence on how embodied GAI agents within immersive environments like Mixed Reality (MR) influence the cognitive and socio‐emotional interactions for learning and regulation. What this paper adds Provides first empirical insights into cognitive and socio‐emotional interaction patterns between learners and embodied GAI agents in MR environments. Identifies two distinct interaction patterns: AISQ type (structured, guided, supportive) and LII type (inquiry‐driven, exploratory, engaging), demonstrating how these patterns influence collaborative learning dynamics. Shows that both interaction types facilitate meaningful cognitive engagement, moving beyond superficial content reproduction commonly associated with GAI interactions. Implications for practice and/or policy Insights from the identified interaction patterns can inform the design of teaching strategies that effectively integrate embodied GAI agents to enhance both cognitive and socio‐emotional engagement. Findings can guide the development of AI‐based educational tools that capitalise on the capabilities of embodied GAI agents, supporting a balance between structured guidance and exploratory learning. Highlights the need for ethical considerations in adopting embodied GAI agents, particularly regarding the human‐like realism of these agents and potential impacts on learner dependency and interaction norms.

No abstract available

In multi-robot exploration, a team of mobile robot is tasked with efficiently mapping an unknown environments. While most exploration planners assume omnidirectional sensors like LiDAR, this is impractical for small robots such as drones, where lightweight, directional sensors like cameras may be the only option due to payload constraints. These sensors have a constrained field-of-view (FoV), which adds complexity to the exploration problem, requiring not only optimal robot positioning but also sensor orientation during movement. In this work, we propose MARVEL, a neural framework that leverages graph attention networks, together with novel frontiers and orientation features fusion technique, to develop a collaborative, decentralized policy using multi-agent reinforcement learning (MARL) for robots with constrained FoV. To handle the large action space of viewpoints planning, we further introduce a novel information-driven action pruning strategy. MARVEL improves multi-robot coordination and decision-making in challenging large-scale indoor environments, while adapting to various team sizes and sensor configurations (i.e., FoV and sensor range) without additional training. Our extensive evaluation shows that MARVEL's learned policies exhibit effective coordinated behaviors, outperforming state-of-the-art exploration planners across multiple metrics. We experimentally demonstrate MARVEL's generalizability in large-scale environments, of up to 90 m by 90 m, and validate its practical applicability through successful deployment on a team of real drone hardware.

Traffic congestion remains a major challenge for modern urban transportation, diminishing both efficiency and quality of life. While autonomous driving technologies and reinforcement learning (RL) have shown promise for improving traffic control, most prior work has focused on small-scale networks or isolated intersections. Large-scale mixed traffic control, involving both human-driven and robotic vehicles, remains underexplored. In this study, we propose a decentralized multi-agent reinforcement learning framework for managing large-scale mixed traffic networks, where intersections are controlled either by traditional traffic signals or by robotic vehicles. We evaluate our approach on a real-world network of 14 intersections in Colorado Springs, Colorado, USA, using average vehicle waiting time as the primary measure of traffic efficiency. We are exploring a problem that has not been sufficiently addressed: Is large-scale Multi-Agent Traffic Control (MTC) still feasible when facing time-varying Origin-Destination (OD) patterns?

Traffic congestion remains a significant challenge in modern urban networks. Autonomous driving technologies have emerged as a potential solution. Among traffic control methods, reinforcement learning has shown superior performance over traffic signals in various scenarios. However, prior research has largely focused on small-scale networks or isolated intersections, leaving large-scale mixed traffic control largely unexplored. This study presents the first attempt to use decentralized multi-agent reinforcement learning for large-scale mixed traffic control in which some intersections are managed by traffic signals and others by robot vehicles. Evaluating a real-world network in Colorado Springs, CO, USA with 14 intersections, we measure traffic efficiency via average waiting time of vehicles at intersections and the number of vehicles reaching their destinations within a time window (i.e., throughput). At 80% RV penetration rate, our method reduces waiting time from 6.17 s to 5.09 s and increases throughput from 454 vehicles per 500 seconds to 493 vehicles per 500 seconds, outperforming the baseline of fully signalized intersections. These findings suggest that integrating reinforcement learning-based control large-scale traffic can improve overall efficiency and may inform future urban planning strategies.

No abstract available

Recently, existing computation offloading methods have provided extremely low service latency for mobile users (MUs) in multi-access edge computing (MEC). However, this remains a challenge in large-scale mixed cooperative-competitive MUs heterogeneous MEC environments. Moreover, existing methods focus more on all offloaded tasks handled by static resource allocation MEC servers (ESs) within a time interval, ignoring on-demand requirements of heterogeneous tasks, resulting in many tasks being dropped or wasting resources, especially for latency-sensitive tasks. To address these issues, we present a decentralized computation offloading solution based on the Attention-weighted Recurrent Multi-Agent Actor-Critic (ARMAAC). First, we design a recurrent actor-critic framework to assist MU agents in remembering historical resource allocation information of ESs to better understand the future state of ESs, especially in dynamic resource allocation. Second, an attention mechanism is introduced to compress the joint observation space dimension of all MUs agent to adapt to large-scale MUs. Finally, the actor-critic framework with double centralized critics and Dueling network is redesigned considering the instability and convergence difficulties caused by the sensitive relationship between the actor and critic networks. The experiments show that ARMAAC improves task completion rates and reduces average system cost by 11.01%<inline-formula><tex-math notation="LaTeX">$\sim$</tex-math><alternatives><mml:math><mml:mo>∼</mml:mo></mml:math><inline-graphic xlink:href="gao-ieq1-3141080.gif"/></alternatives></inline-formula>14.03% and 10.45%<inline-formula><tex-math notation="LaTeX">$\sim$</tex-math><alternatives><mml:math><mml:mo>∼</mml:mo></mml:math><inline-graphic xlink:href="gao-ieq2-3141080.gif"/></alternatives></inline-formula>15.56% compared with baselines.

With the exponential growth of mobile users, ensuring high-quality network coverage has become paramount. Large-scale mobile networks consist of numerous base stations (BSs), each with adjustable parameters such as angles and beam widths. Automatically optimizing network coverage can be difficult due to environmental factors and the interdependence of the adjustable parameters. Due to the inherent uncertainties and unpredictable nature of large-scale wireless networks, traditional methods such as heuristics and meta-heuristics lack the adaptability and scalability required to cope with their dynamic environment. To address these challenges, we propose utilizing digital twin and reinforcement learning (RL) techniques within mobile networks characterized by multiple collaborating agents. We initially introduce DT-SimNet, a digital twin-enabled mobile network simulator to facilitate optimization evaluation. DT-SimNet can efficiently simulate communication behaviors of network elements within a complex environment while revealing user mobility patterns. Moreover, to address challenges arising from multifaceted relationships among users, BSs, and the parameters across BSs, we introduce an innovative strategy named Optimized Multi-Agent Proximal Policy Optimization with Self-supervised Prediction (OMAPPO-SSP). Compared to MAPPO, which leads to limited applicability and inferior performance due to the dynamic characteristics of 5G networks, this approach leverages network structure optimization and a self-supervised prediction mechanism, employing multi-agent reinforcement learning (MARL) principles to enhance efficiency. By harnessing collaborative neural networks, OMAPPO-SSP facilitates the explicit learning of behavioral interactions among all BSs, enabling effective decision-making in environments characterized by intricate spatial relationships, dynamic user behaviors, and diverse interactions. Extensive experiments are conducted to validate the efficiency and effectiveness of the OMAPPO-SSP. Within the target area, OMAPPO-SSP achieves a coverage ratio of 94.66% and an average throughput of 89746 bits per second (bps), demonstrating significant improvements compared to competing methods.

In this paper, the real-time optimal transmission power control problem is investigated for large-scale mobile wireless sensor networks (MWSN). Controlling large-scale MWSN has two novel challenges, i.e., 1) increasing navigation complexity due to a large number of mobile wireless sensors, and 2) limited energy prohibits peer-to-peer communication between large-scale mobile sensors. To overcome these challenges, the novel mean field game theory is adopted and integrated along with the emerging decentralized reinforcement learning technique. Specifically, the optimal transmission control problem and the optimal navigation problem are formulated as mean field games with two objectives. Then, a novel Actor-Critic-Mass multi-agent reinforcement learning algorithm is developed to learn the decentralized optimal transmission power control and motion control. To learn the decentralized optimal navigation and transmission power control policies, the coupled Hamiltonian-Jacobian-Bellman (HJB) and Fokker-Planck-Kolmogorov (FPK) equations are derived in mean field game formulation. The learned decentralized policies can be guaranteed to converge close to the optimal value, i.e., the Nash Equilibrium, even with large-scale MWSNs in uncertain environments. Finally, the numerical simulations have been provided to demonstrate the effectiveness of the proposed design.

With the rapid advancement of communication technology and the exponential growth of mobile users, improving network coverage quality and throughput has become increasingly important. In particular, large-scale Base Station (BS) cooperative optimization has become a highly significant topic. BSs can adjust various parameters for high-quality communication, but automating this optimization remains challenging due to environmental sensitivity and interdependencies. Traditional methods for network optimization are constrained by the intricate nature of real-world environments. Further, Reinforcement Learning (RL) techniques, which are effective for configuration policies, encounter difficulties in intricate, high-dimensional wireless communication networks, especially in multi-agent cooperative optimization. To overcome these challenges, this article proposes the Enhanced Multi-Agent Proximal Policy Optimization (EMAPPO), which utilizes the capabilities of the UNet network to extract multi-spatial relationships among a massive number of network elements and employs the DiffPool network to efficiently depict the impact of large-scale action coordination among massive agents on coverage performance. To facilitate evaluation in communication optimization, we further introduce a high-fidelity digital twin-driven mobile network. Extensive experiments validate the effectiveness and superior performance of EMAPPO by utilizing the network digital twin. The results demonstrate significant improvements in signal coverage rate and network throughput compared to the competing methods.

The use of multi-agent reinforcement learning (MARL) methods in coordinating traffic lights (CTL) has become increasingly popular, treating each intersection as an agent. However, existing MARL approaches either treat each agent absolutely homogeneous, i.e., same network and parameter for each agent, or treat each agent completely heterogeneous, i.e., different networks and parameters for each agent. This creates a difficult balance between accuracy and complexity, especially in large-scale CTL. To address this challenge, we propose a grouped MARL method named GPLight. We first mine the similarity between agent environment considering both real-time traffic flow and static fine-grained road topology. Then we propose two loss functions to maintain a learnable and dynamic clustering, one that uses mutual information estimation for better stability, and the other that maximizes separability between groups. Finally, GPLight enforces the agents in a group to share the same network and parameters. This approach reduces complexity by promoting cooperation within the same group of agents while reflecting differences between groups to ensure accuracy. To verify the effectiveness of our method, we conduct experiments on both synthetic and real-world datasets, with up to 1,089 intersections. Compared with state-of-the-art methods, experiment results demonstrate the superiority of our proposed method, especially in large-scale CTL.

This letter develops a novel multi-agent deep reinforcement learning (MADRL)-based local control method that can achieve coordinated scheduling of large-scale PV inverters using local information. This is achieved by the development of a system state inference-aided actor structure for each agent and implementation of random sequential updating within centralized-training-decentralized-execution framework. To enhance the coordination between agents utilizing local observation, a state latent inductive reasoning-based composite loss is further designed for the optimization of the inference models. Simulation tests on IEEE 123-node network demonstrate the superiority of the developed local control method when there is a large number of PV inverters.

Traffic signal control (TSC) is a challenging problem within intelligent transportation systems and has been tackled using multi-agent reinforcement learning (MARL). While centralized approaches are often infeasible for large-scale TSC problems, decentralized approaches provide scalability but introduce new challenges, such as partial observability. Communication plays a critical role in decentralized MARL, as agents must learn to exchange information using messages to better understand the system and achieve effective coordination. Deep MARL has been used to enable inter-agent communication by learning communication protocols in a differentiable manner. However, many deep MARL communication frameworks proposed for TSC allow agents to communicate with all other agents at all times, which can add to the existing noise in the system and degrade overall performance. In this study, we propose a communication-based MARL framework for large-scale TSC. Our framework allows each agent to learn a communication policy that dictates “which” part of the message is sent “to whom”. In essence, our framework enables agents to selectively choose the recipients of their messages and exchange variable length messages with them. This results in a decentralized and flexible communication mechanism in which agents can effectively use the communication channel only when necessary. We designed two networks, a synthetic $4 \times 4$ grid network and a real-world network based on the Pasubio neighborhood in Bologna. Our framework achieved the lowest network congestion compared to related methods, with agents utilizing $\sim 47-65 \%$ of the communication channel. Ablation studies further demonstrated the effectiveness of the communication policies learned within our framework.

Traffic signal control is an emerging application scenario for reinforcement learning. Besides being as an important problem that affects people's daily life in commuting, traffic signal control poses its unique challenges for reinforcement learning in terms of adapting to dynamic traffic environment and coordinating thousands of agents including vehicles and pedestrians. A key factor in the success of modern reinforcement learning relies on a good simulator to generate a large number of data samples for learning. The most commonly used open-source traffic simulator SUMO is, however, not scalable to large road network and large traffic flow, which hinders the study of reinforcement learning on traffic scenarios. This motivates us to create a new traffic simulator CityFlow with fundamentally optimized data structures and efficient algorithms. CityFlow can support flexible definitions for road network and traffic flow based on synthetic and real-world data. It also provides user-friendly interface for reinforcement learning. Most importantly, CityFlow is more than twenty times faster than SUMO and is capable of supporting city-wide traffic simulation with an interactive render for monitoring. Besides traffic signal control, CityFlow could serve as the base for other transportation studies and can create new possibilities to test machine learning methods in the intelligent transportation domain.

Recent advancements in reinforcement learning have witnessed remarkable achievements by intelligent agents ranging from game-playing to industrial applications. Of particular interest is the area of multi-agent reinforcement learning (MARL), which holds significant potential for real-world scenarios. However, typical MARL methods are limited in their ability to handle tens of agents, leaving scenarios with up to hundreds or even thousands of agents almost unexplored. The scaling up of the number of agents presents two primary challenges: (1) agent-agent interactions are crucial in multi-agent systems while the number of interactions grows quadratically with the number of agents, resulting in substantial computational complexity and difficulty in strategies-learning; (2) the strengths of interactions among agents exhibit variations both across agents and over time, making it difficult to precisely model such interactions. In this paper, we propose a novel approach named Graph Attention Mean Field (GAT-MF). By converting agent-agent interactions into interactions between each agent and a weighted mean field, we achieve a substantial reduction in computational complexity. The proposed method offers a precise modeling of interaction dynamics with mathematical proofs of its correctness. Additionally, we design a graph attention mechanism to automatically capture the diverse and time-varying strengths of interactions, ensuring an accurate representation of agent interactions. Through extensive experimentation conducted in both manual and real-world scenarios involving over 3000 agents, we validate the efficacy of our method. The results demonstrate that our method outperforms the best baseline method with a remarkable improvement of 42.7%. Furthermore, our method saves 86.4% training time and 19.2% GPU memory compared to the best baseline method. For reproducibility, our source codes and data are available at https://github.com/tsinghua-fib-lab/Large-Scale-MARL-GATMF.

With the emerging connected-vehicle technologies and smart roadways, the need for intelligent adaptive traffic signal controls (ATSC) is more than ever before. This paper first proposes an Accumulated Exponentially Weighted Waiting Time-based Adaptive Traffic Signal Control (AEWWT-ATSC) model to calculate priorities of roadways for signal scheduling. As the size of the traffic network grows, it adds great complexities and challenges to computational efficiencies. Considering this, we propose a novel Distributed Multi-agent Reinforcement Learning (DMARL) with a graph decomposition approach for large-scale ATSC problems. The decomposition clusters intersections by the level of connectivity (LoC), defined by the average residual capacities (ARC) between connected intersections, enabling us to train subgraphs instead of the entire network in a synchronized way. The problem is formulated as a Markov Decision Process (MDP), and the Double Dueling Deep Q Network with Prioritized Experience Replay is utilized to solve it. Under the optimal policy, the agents can select the optimal signal durations to minimize the waiting time and queue size. In evaluation, we show the superiority of the AEWWT-ATSC based RL methods in different densities and demonstrate the DMARL with a graph decomposition approach on a large graph in Manhattan, NYC. The approach is generic and can be extended to various types of use cases.

No abstract available

Robotic sortation centers use mobile robots to sort packages by their destinations. The destination-to-sort-location (chute) mapping can significantly impact the volume of packages that can be sorted by the sortation floor. In this work, we propose a multi-agent reinforcement learning method to solve large-scale chute mapping problems with hundreds of agents (the destinations). To address the exponential growth of the state-action space, we decompose the joint action-value function as the sum of local action-value functions associated with the individual agents. To incorporate robot congestion effects on the rates at which packages are sorted, we couple the local action-value functions through the states of destinations mapped to nearby chutes on the sortation floor. We show that our proposed framework can solve large chute mapping problems and outperforms static or reactive policies that are commonly used in practice in robotic sortation facilities.

Variable speed limit (VSL) control has emerged as a promising traffic management strategy for enhancing safety and mobility. In this study, we introduce a multi-agent reinforcement learning framework for implementing a large-scale VSL system to address recurring congestion in transportation corridors. The VSL control problem is modeled as a Markov game, using only data widely available on freeways. By employing parameter sharing among all VSL agents, the proposed algorithm can efficiently scale to cover extensive corridors. The agents are trained using a reward structure that incorporates adaptability, safety, mobility, and penalty terms; enabling agents to learn a coordinated policy that effectively reduces spatial speed variations while minimizing the impact on mobility. Our findings reveal that the proposed algorithm leads to a significant reduction in speed variation, which holds the potential to reduce incidents. Furthermore, the proposed approach performs satisfactorily under varying traffic demand and compliance rates.

No abstract available

In this paper, we investigate the uplink transmit power optimization problem in cell-free (CF) extremely large-scale multiple-input multiple-output (XL-MIMO) systems. Instead of applying the traditional methods, we propose two signal processing architectures: the centralized training and centralized execution with fuzzy logic as well as the centralized training and decentralized execution with fuzzy logic, respectively, which adopt the amalgamation of multi-agent reinforcement learning (MARL) and fuzzy logic to solve the design problem of power control for the maximization of the system spectral efficiency (SE). Furthermore, the uplink performance of the system adopting maximum ratio (MR) combining and local minimum mean-squared error (L-MMSE) combining is evaluated. Our results show that the proposed methods with fuzzy logic outperform the conventional MARL-based method and signal processing methods in terms of computational complexity. Also, the SE performance under MR combining is even better than that of the conventional MARL-based method.

Reinforcement learning (RL) is a promising data-driven approach for adaptive traffic signal control (ATSC) in complex urban traffic networks, and deep neural networks further enhance its learning power. However, the centralized RL is infeasible for large-scale ATSC due to the extremely high dimension of the joint action space. The multi-agent RL (MARL) overcomes the scalability issue by distributing the global control to each local RL agent, but it introduces new challenges: now, the environment becomes partially observable from the viewpoint of each local agent due to limited communication among agents. Most existing studies in MARL focus on designing efficient communication and coordination among traditional Q-learning agents. This paper presents, for the first time, a fully scalable and decentralized MARL algorithm for the state-of-the-art deep RL agent, advantage actor critic (A2C), within the context of ATSC. In particular, two methods are proposed to stabilize the learning procedure, by improving the observability and reducing the learning difficulty of each local agent. The proposed multi-agent A2C is compared against independent A2C and independent Q-learning algorithms, in both a large synthetic traffic grid and a large real-world traffic network of Monaco city, under simulated peak-hour traffic dynamics. The results demonstrate its optimality, robustness, and sample efficiency over the other state-of-the-art decentralized MARL algorithms.