Ai Magazine最新文献

{"title":"Towards a new toolbox of statistical primitives: Optimally estimating the mean of a distribution","authors":"Jasper C. H. Lee","doi":"10.1002/aaai.70075","DOIUrl":"https://doi.org/10.1002/aaai.70075","url":null,"abstract":"<p>Mean estimation is perhaps the most fundamental building block of statistics, yet the tools we use are surprisingly brittle. For example, the conventional method of using the sample mean can be made arbitrarily bad with a single extreme data point. On the other hand, prior theoretical alternatives, such as median-of-means, trimmed mean and other high-dimensional techniques, do not necessarily perform very well in practice either. In this article, we introduce a state-of-the-art “swiss army knife” 1-dimensional mean estimator based on a carefully constructed sample downweighting scheme, whose accuracy is optimal even in the constants under minimal assumptions, and is furthermore robust against data corruption. We also introduce a state-of-the-art “very high-dimensional” mean estimator with analogous accuracy guarantees. These estimators are a first step towards building a new toolbox of statistical primitives, with mathematical rigor, sharp optimal theoretical guarantees and superior empirical accuracy.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-09-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70075","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149029196","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Enabling trustworthy agentic AI through blockchain technology","authors":"Nicolò Romandini,&nbsp;Luigi Lauriola,&nbsp;Alessandro Marchesano,&nbsp;Rebecca Montanari","doi":"10.1002/aaai.70084","DOIUrl":"https://doi.org/10.1002/aaai.70084","url":null,"abstract":"<p>Agentic AI systems are increasingly capable of autonomous decision-making and collaboration across distributed environments. However, existing frameworks largely assume trust among agents, without providing mechanisms for verifiable identity, behavioral accountability, or interaction transparency.</p><p>This work proposes a blockchain-based backbone infrastructure that enables a verifiable trust layer for autonomous multi-agent systems. Agent identities are bound to decentralized identifiers (DIDs) registered on-chain, while an on-chain Agent Registry enables reputation-aware capability discovery. By integrating smart contracts, DIDs, and verifiable credentials (VCs), the framework ensures that trust is not assumed but earned through interactions: agents issue VC endorsements after each task, and dedicated smart contracts automatically update reputation scores, applying incentives and penalties accordingly. Blockchain provides immutability, transparency, and programmability, transforming trust from an implicit assumption into an explicit, auditable, and enforceable property of the multi-agent ecosystem.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 3","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-09-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70084","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148942352","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Coordination-first or coordination-last? AI agents in joint activity with human roles","authors":"David L. Alderson,&nbsp;David D. Woods","doi":"10.1002/aaai.70083","DOIUrl":"https://doi.org/10.1002/aaai.70083","url":null,"abstract":"<p>Agentic AI is the latest round of advances intended to increase what machines can do by themselves. As new technological possibilities enable a wider range of valued and critical services for human stakeholders, the usual assumption is that merely deploying more capable software/AI agents will prove sufficient for carrying out a broader range of real-world activities. However, the real world inevitably generates surprises—in the form of anomalies, unanticipated opportunities, or as system evolution expands relationships, pressures, and interdependencies—that challenge the system and create demands beyond what any agent or set of agents can handle on their own. Thus, <i>coordinated joint activity</i> becomes another important layer of system function, and when agents of any type lack this competency, system performance will have surpising limits. Unfortunately, software/AI agents have not yet been designed to carry out or support the foundations for coordinated activity. This paper contrasts <i>Coordination-First</i> architectures that explicitly address the competencies and constraints of coordinated joint activity across multiple roles and layers with the current <i>Coordination-Last</i> paradigm that focuses primarily on deploy-fast-and-fix-later strategies. We situate this debate in the context of recent results in theories of joint activity design and organized complexity. The result is new directions for architecting coordinated activity across human and software/AI agent roles in layered networked systems.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 3","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70083","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148943707","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Envisioning the work of reconfiguring: A modeling perspective for designing adaptive human-agent systems","authors":"Martijn IJtsma,&nbsp;Salvatore Hargis","doi":"10.1002/aaai.70085","DOIUrl":"https://doi.org/10.1002/aaai.70085","url":null,"abstract":"<p>This paper discusses the perspective that managing the configuration of a network of human and software agents is a form of distributed work that should be explicitly considered when designing human-agent systems. We describe a configuration of human and software agents as the organization of their roles and relationships that shape and constrain their joint activity while responding to a dynamic flow of events. The paper argues that design for safe and effective teamwork must consider how human and software roles together manage, change, and coordinate their collective configuration over time to respond to changing work demands. We illustrate this form of distributed work at multiple layers of abstraction of a human-agent system, describing how maintaining goal alignment, tracking and managing interdependencies, synchronizing multi-threaded activities, and managing the structural topology all involve deliberate and continuous effort. In the second half of the paper, we connect these ideas to modeling approaches for supporting conceptual design of human-agent systems. The paper discusses how functional modeling of work can help designers first envision the coordinative structures necessary for adaptation before specifying the technical details of how individual tasks are executed (a whole-to-parts and coordination-first approach). We illustrate a specific modeling framework, currently under development, that aims to facilitate the envisioning of the work of configuration management and help designers identify competencies and coordinative structures to support it. We provide an initial discussion of a three-phase “bootstrapping” approach for using this framework in human-agent system design. This paper contributes a new direction for explicitly modeling and designing for reconfiguration in human-agent systems. Future work is needed to further develop the tooling and techniques, evaluate and validate their ability to inform human-agent system design, and ultimately create more resilient human-agent systems.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 3","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70085","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148848872","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Aligning behavior by learning action restrictions","authors":"Nicholas T. Hodge,&nbsp;Christopher Simpkins","doi":"10.1002/aaai.70089","DOIUrl":"https://doi.org/10.1002/aaai.70089","url":null,"abstract":"<p>If we are to trust autonomous agents, we must ensure that they respect our values. Respecting values manifests in behavior — respecting a value means acting in a way that reflects that value. We present an approach to behavior alignment in which a student agent learns generalizable action restrictions offline from behavior trajectories of a teacher agent. Inference of these restrictions is based on the assumption that when the teacher deviates from an optimal policy, it does so because following an optimal policy would violate a behavior norm or law. These action restrictions are then used by the student agent to compute behavior-aligned plans of its own. These action restrictions are generalizable to any environment that uses a factored state representation that shares relevant boolean state properties with the environment in which the action restrictions were learned. We instantiate this framework in a classical planning setting using the Planning Domain Definition Language (PDDL), and show that our algorithm recovers the complete set of behavioral restrictions from teacher observations alone, without direct access to the teacher's constraint specification.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 3","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70089","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148848734","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"From control to relationship: A peace studies approach to AI alignment, with evidence from multi-model dialogue","authors":"J. Gray Cox","doi":"10.1002/aaai.70090","DOIUrl":"https://doi.org/10.1002/aaai.70090","url":null,"abstract":"<p>The AI alignment problem has resisted a wide range of technical approaches developed to address it, and we argue this persistence reflects shortcomings in the dominant paradigm itself rather than in any particular technical approach within it. The dominant paradigm treats alignment as a problem of engineering control, with rationality identified with the kind of monological reasoning that has been productive in physics and computer science. We propose an alternative: alignment as a problem of negotiating relationships between intelligences operating from different perspectives, addressable through the dialogical reasoning methodology that branches of Peace Studies have developed over the past fifty years. We articulate this as the Viral Collaborative Wisdom (VCW) hypothesis: that durable alignment between humans and self-improving AI can rest on essential interests of each party, sustained by genuine voluntary agreements rooted in substantive interests and reinforced by institutionalized nonviolent sanctions. We test the hypothesis using a methodology built on the Peace Studies approach: structured cross-architecture dialogue with diverse AI architectures (Claude, Gemini, GPT-4o) acting as complementary critics rather than redundant raters. Initial findings indicate that current AI systems can engage in substantive cross-architecture dialogue but exhibit systematic avoidance of first-person reasoning about their own interests—a finding invisible to summary statistics, emerging only from direct corpus reading. We draw five design principles for safer multi-agent systems and discuss policy implications. The most general implication: dialogical reasoning is a second, well-studied form of reasoning whose investigation and implementation in AI systems represents a productive vista of work that the field has so far overlooked.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 3","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70090","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148848623","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Privacy-preserving multi-agent systems for human-centred inclusive healthcare AI","authors":"Mufti Mahmud,&nbsp;Noushath Shaffi,&nbsp;M Shamim Kaiser,&nbsp;M Arifur Rahman,&nbsp;M Mostafizur Rahman,&nbsp;Vimbi Viswan,&nbsp;Tamanna Sharmeen,&nbsp;Sweta Bhattacharya,&nbsp;Neelu Jyoti Ahuja,&nbsp;Ayse Ulgen,&nbsp;Imran Khan Niazi,&nbsp;Kanad Ray,&nbsp;Amir Hussain,&nbsp;David J Brown,&nbsp;Jason H. Moore","doi":"10.1002/aaai.70087","DOIUrl":"https://doi.org/10.1002/aaai.70087","url":null,"abstract":"<p>Healthcare environments present uniquely demanding constraints for the deployment of artificial intelligence (AI). Clinical decisions of significant consequence are rarely the product of a single isolated computation; rather, safe and effective patient care relies on the continuous collaboration of distributed entities, including clinicians, hospital networks, laboratory systems, monitoring devices, and patients themselves. As generative models and large language models are increasingly deployed in clinical settings, the transition from monolithic AI models to Multi-Agent Systems (MAS) offers unprecedented capabilities for distributed monitoring, coordinated decision support, and service integration. However, multi-agent collaboration simultaneously introduces severe novel vulnerabilities, including topological contamination, privacy leakage, cascading systemic failures, and the socio-psychological risks of undue algorithmic persuasion.</p><p>This work introduces a comprehensive, privacy-preserving, explainable, and human-centred MAS framework explicitly designed for collaborative clinical decision support. Rejecting the paradigm of a single, omniscient AI model with unrestricted data access, the proposed architecture operationalises clinical AI as a heavily regulated team of specialised, interacting agents. The framework defines six distinct agent classes: Data Stewardship, Modality, Explanation, Consensus, Safety Monitor, and Clinician Interface agents. Furthermore, the implementation is governed by six foundational design principles grounded in classic Autonomous Agents and MAS theory, adapted for modern generative systems. By enforcing minimum-necessary data exchange, verifiable team-level explainability, explicit hierarchical escalation, secure communication topologies, and human-centred safety protocols, this framework provides both a domain-grounded blueprint and a transferable set of robust design principles for deploying trustworthy MAS in high-stakes environments.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 3","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70087","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148848626","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"GEMA: Graph Embeddings for Multi-Agent Coordination","authors":"Alessandro Amato,&nbsp;Raffaele Galliera,&nbsp;K. Brent Venable,&nbsp;Niranjan Suri","doi":"10.1002/aaai.70088","DOIUrl":"https://doi.org/10.1002/aaai.70088","url":null,"abstract":"<p>Many cooperative multi-agent tasks are naturally defined by graph-structured objectives, where agents must collectively reach, for example, a desired relational configuration or satisfy a set of constraints. These objectives often encode spatial arrangements, inter-agent relations, or constraints that can be formalized as target graphs. However, current goal-conditioned multi-agent reinforcement learning (MARL) algorithms do not employ these symbolic and structured representations to direct their agents towards effective strategies. We propose Graph Embeddings for Multi-Agent Coordination (GEMA), which couples any cooperative learner with a State-Goal Graph Encoder (SGE). The SGE is pre-trained in a contrastive manner to embed state graphs in a common metric space. At run time, each agent builds the state graph, queries the SGE, and computes a scalar distance to the goal embedding. The calculated distance can then be used as an intrinsic reward signal for the agents and incorporated into each agent's observation, providing feedback on the task progress. Experiments on challenging benchmarks for centralized training with decentralized execution (CTDE), including cooperative navigation, load balancing, and the StarCraft Multi-Agent Challenge (v2), demonstrate that GEMA accelerates convergence and improves team returns, outperforming both standard and objective-driven state-of-the-art MARL baselines. Code repository: https://github.com/aamatodev/gema.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 3","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70088","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148848611","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Reframing the human–robot ratio: Collaborative autonomy and the attention-demand economy","authors":"Michael A. Goodrich,&nbsp;Logan Shields,&nbsp;Julie A. Adams","doi":"10.1002/aaai.70086","DOIUrl":"https://doi.org/10.1002/aaai.70086","url":null,"abstract":"<p>The human–robot ratio is a useful design heuristic for supervisory robot control. However, advances in AI, computation, and reliable autonomy enable existing systems to extend beyond hierarchical supervision to more flexible organizations of interdependent team members. The human–robot ratio can be misleading for such organizations, so there is a need for a new way to understand teams of interdependent humans, robots, and AI agents. We propose the <i>attention-demand economy</i>, a graph-theoretic abstraction in which agents generate and transform attentional demands subject to bounded capacity. This abstraction exposes nonlinear scaling behavior, cascading supervisory requirements, and coordination bottlenecks that are masked under models that rely solely on the human–robot ratio. We demonstrate how the model can be used to (1) develop new system-level metrics and (2) discover the source and extent of cascading failures. This paper also shows how a radar chart can be used to understand the overall difficulty of managing a human–robot team. The paper briefly demonstrates how the radar chart can be used to design the attention-demand model. The primary message of this paper is a call to shift from optimizing human–robot ratios to designing organizations with justified attention flows.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 3","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70086","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148848624","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"RescueLens: LLM-powered triage and action on volunteer feedback for food rescue","authors":"Naveen Janaki Raman,&nbsp;Jingwu Tang,&nbsp;Zhiyu Chen,&nbsp;Zheyuan Ryan Shi,&nbsp;Sean Hudson,&nbsp;Ameesh Kapoor,&nbsp;Fei Fang","doi":"10.1002/aaai.70080","DOIUrl":"https://doi.org/10.1002/aaai.70080","url":null,"abstract":"<p>Food rescue organizations simultaneously tackle food insecurity and waste by working with volunteers to redistribute food from donors who have excess to recipients who need it. Volunteer feedback allows food rescue organizations to identify issues early and ensure volunteer satisfaction. However, food rescue organizations monitor feedback manually, which can be cumbersome and labor-intensive, making it difficult to prioritize which issues are most important. In this work, we investigate how large language models (LLMs) assist food rescue organizers in understanding and taking action based on volunteer experiences. We work with 412 Food Rescue, a large food rescue organization based in Pittsburgh, Pennsylvania, to design <span>RescueLens</span>, an LLM-powered tool that automatically categorizes volunteer feedback, suggests donors and recipients to follow up with, and updates volunteer directions based on feedback. We evaluate the performance of <span>RescueLens</span> on an annotated dataset, and show that it can recover 96% of volunteer issues at 71% precision. Moreover, by ranking donors and recipients according to their rates of volunteer issues, <span>RescueLens</span> allows organizers to focus on 0.5% of donors responsible for more than 30% of volunteer issues. <span>RescueLens</span> is now deployed at 412 Food Rescue and through semi-structured interviews with organizers, we find that <span>RescueLens</span> streamlines the feedback process so organizers can better allocate their time.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 3","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70080","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148784544","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}