Ai Magazine最新文献

{"title":"Tools are all you need","authors":"Henry Kautz","doi":"10.1002/aaai.70106","DOIUrl":"https://doi.org/10.1002/aaai.70106","url":null,"abstract":"<p>Even while AI based on large language models makes enormous advances, many AI researchers and pundits argue that it LLMs are an inadequate basis for AI because they do not support exact logical and probabilistic reasoning. I argue in response that LLMs can and are using tools to perform such normative reasoning, and thus are in face already examples of general intelligence.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70106","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149205920","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":"AI, intelligence, and children","authors":"Moshe Y. Vardi","doi":"10.1002/aaai.70111","DOIUrl":"https://doi.org/10.1002/aaai.70111","url":null,"abstract":"&lt;p&gt;The nature of intelligence is a major scientific question. Daniel Kahneman made major contributions to the study of human intelligence. A related question is whether machines can be considered to be intelligent. This question goes back to the 19&lt;sup&gt;th&lt;/sup&gt; Century. “&lt;i&gt;Precisely how much the business of thinking a machine could possibly be made to perform, and what part of it must be left to the living mind is a question not without conceivable practical importance&lt;/i&gt;,” wrote the American philosopher Charles S. Pierce in 1887. In his &lt;i&gt;Computing Machinery and Intelligence&lt;/i&gt; (1950), Alan M. Turing proposed the “Imitation Game”, now known as the “Turing Test”. In that test, a human evaluator attempts to distinguish between a human and a machine by means of a (technology-mediated) conversation. An intelligent machine should be able to converse indistinguishably from a human.&lt;/p&gt;&lt;p&gt;The Turing Test has been a subject of intense philosophical debate since its publication. Its most famous criticism has been by the philosopher John Searle in his Chinese Room thought experiment. But this essay focuses on an aspect of the Turing Test that has generally escaped attention. In 1948, two years before his “Computing Machinery and Intelligence,” Turing wrote a technical report for the National Physical Laboratory in Manchester. The report, titled “Intelligent Machinery,” was effectively the first manifesto of Artificial Intelligence. In that report Turing already proposed the Imitation Game, but the proposed interaction between the evaluator and the human/machine was via a chess game, rather than a conversation. In 1950, Turing changed the test from chess playing to conversing.&lt;/p&gt;&lt;p&gt;But the key difference between the 1948 paper and the 1950 paper was not in the details of the test. Rather, in the 1948 Turing did not attempt to define intelligence. Instead, he referred to intelligence as “an emotional concept”. He wrote: “The extent to which we regard something as behaving in an intelligent manner as much by our own state of mind and training as by the properties of the object under consideration.” In other words, the Turing Test, as envisioned by Turing in 1948, is a test of whether we could build machines that &lt;b&gt;&lt;i&gt;appear to humans&lt;/i&gt;&lt;/b&gt; to be intelligent. Clearly, that test has been passed by ChatGPT when it was launched in late 2022. ChatGPT and other LLMs clearly appear to be intelligent. The rest of this essay argues that the availability of “apparently intelligent” machinery has profound implications for the mental health of young people.&lt;/p&gt;&lt;p&gt;To start, we must recognize that a crisis has been brewing in the U.S. over the past few years—the youth mental health crisis. In 2021, the U.S. Surgeon General issued an advisory report titled &lt;i&gt;Protecting Youth Mental Health&lt;/i&gt;. “The challenges today's generation of young people face are unprecedented and uniquely hard to navigate,” wrote the Surgeon General, adding, “the effect these challenges hav","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70111","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149206089","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":"Fragmented control, collective risk: multi-agent AI and the coming governance crisis","authors":"Caryn Lusinchi","doi":"10.1002/aaai.70105","DOIUrl":"https://doi.org/10.1002/aaai.70105","url":null,"abstract":"<p>Multi-agent artificial intelligence (MAAI) systems are increasingly deployed across economic and digital infrastructures. Existing regulatory frameworks, including the EU AI Act and related initiatives, largely focus on individual AI systems and are therefore poorly suited to govern risks that arise from agent interaction. This paper argues that multi-agent AI introduces distinct governance challenges, including emergent behavior, tacit coordination, cascading systemic risks, and difficulties in assigning accountability. Traditional risk management approaches, which rely on static and component-based analysis, are inadequate for adaptive, networked systems operating across jurisdictions.</p><p>Through a comparative review of policy developments in the European Union, United States, and Asia-Pacific region, the paper identifies a growing governance gap, alongside early regulatory experimentation such as China's emerging standards for intelligent agents. The paper concludes that effective oversight will require a shift from model-centric regulation toward system-level and adaptive governance frameworks capable of addressing the collective dynamics of multi-agent AI systems.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70105","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149205670","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":"How to make System 1 LLMs show System 2 competencies?","authors":"Subbarao Kambhampati","doi":"10.1002/aaai.70113","DOIUrl":"https://doi.org/10.1002/aaai.70113","url":null,"abstract":"&lt;p&gt;From their inception, Large Language Models were best understood as a sort of giant external System 1 for the humanity at large (see the right part of the figure). Trained as they are on the collective digital footprints of all of humanity's articulated knowledge, LLMs have shown that their next token prediction mechanism can exhibit impressive System 1 like verbal behaviors. Unlike most AI systems preceding them, which showed “deep and narrow” System 2 competencies on particular tasks (AlphaGo, Deep Blue, etc.), LLMs were the first AI systems that brought forth broad System 1 competence across many (almost all?) tasks.&lt;/p&gt;&lt;p&gt;But why only System 1? How about System 2 behaviors? Despite many claims to the contrary in the early 2022–24 literature, standard LLMs have been shown to fall far short of System 2 behaviors–such as planning and reasoning. Initial excitement based on anecdotal performance of LLMs on reasoning tasks has dissipated to some extent by the later spate of studies questioning the robustness of such behaviors–be it planning (as shown by our own work), simple arithmetic and logic, or general mathematical and abstract benchmarks. Performance of LLMs on reasoning tasks only started improving substantially since fall 2024, when the first so-called “reasoning models” started coming out– OpenAI's o1, DeepSeek's R1 and so forth. The interesting part, however is that these models are still trained auto-regressively and do next token prediction at the inference stage, thus sticking close to the System 1 operation. How then did they acquire reasoning abilities?&lt;/p&gt;&lt;p&gt;Kahneman's account already provides a possible direction–you can “compile” System 2 behaviors into System 1. While humans, with their own System 2, might be essentially memoizing what they do through deliberation once into reflexive reasoning for later use, LLMs depend on external synthetic training data. The currently popular approach can be understood as a way of fine tuning the base model on a huge amount of synthetic data with verified solutions in the so-called “post-training” phase. We will unpack this below.&lt;/p&gt;&lt;p&gt;Let's start with the generally accepted fact that standard LLMs are very good generators of plausible solutions for any task including reasoning tasks. With a generator in hand, we can do a &lt;i&gt;Generate-Test&lt;/i&gt; search to identify correct solution candidates for the reasoning tasks. While generate-test is normally viewed as a highly inefficient search technique, this can change if we have access to a good generator–and LLMs, it turns out, can fill this role rather well as they have been trained on our collective digital footprints. The question is who is going to test the candidates that an LLM generates.&lt;/p&gt;&lt;p&gt;While there is a temptation to make the LLMs themselves be the judges of the correctness of the candidates they generate, research (including ours) suggests that by and large, LLMs may be no better at verification (unless they have been specifically tra","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70113","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149205785","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":"Better together: Multi-Agent system teamwork","authors":"Jeffrey M. Bradshaw,&nbsp;Nancy Cooke,&nbsp;Michael A. Goodrich,&nbsp;Matthew Johnson,&nbsp;Mufti Mahmud,&nbsp;David D. Woods","doi":"10.1002/aaai.70100","DOIUrl":"https://doi.org/10.1002/aaai.70100","url":null,"abstract":"<p>While most present-day multi-agent system (MAS) deployments involve teams of agents within a single institution orchestrated by one or a few actors, some researchers are already beginning to appreciate the promise of more complex ecosystems consisting of hybrid human-agent-robot teams that operate across organizational and geographic boundaries. Importantly, effective teamwork does not emerge merely from making individual agents more capable or from accumulated experience working alongside other agents. Rather, success requires every essential ingredient to be present and planned for; failure requires only one to be missing. In addition to leveraging what is already known about team structure, heterogeneity/diversity, and team processes, research should also account for the diverse task types and forms of interdependence that make up sophisticated, performative joint activity in the real world. Considering the high-stake opportunities and risks of MAS, a group of researchers from around the world have formed the international Multi-Agent System Safety and Teamwork (MASST) initiative. Essays in this this AI Magazine special issue on Multi-Agent System Teamwork—and its companion special issue on Multi-Agent System Safety and Human-Centered Agent Interaction—have been selected from the proceedings of past MASST events. In this article, we outline some of the needed requirements for making MAS “better together.” Innovative research relating to these requirements is grouped under four headings: 1. enabling scalable teamwork; 2. assuring multi-agent alignment; 3. managing the nuanced dynamics of trust and delegation; 4. theory and design of “coordination-first” systems</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70100","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149146434","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":"General-purpose agents in human-machine teams: Can we preserve meaningful human control?","authors":"Karel van den Bosch,&nbsp;Peter-Paul van Maanen,&nbsp;Jurriaan van Diggelen","doi":"10.1002/aaai.70091","DOIUrl":"https://doi.org/10.1002/aaai.70091","url":null,"abstract":"<p>As General-Purpose AI (GPAI) grows more capable of cross-domain reasoning and multi-agent coordination, it reshapes the design space of military human-machine teams and introduces new opportunities for adaptability as well as new risks for Meaningful Human Control (MHC). The paper presents an integrated delegation framework for preserving MHC in GPAI-enabled human-machine teams. It combines play-based task structuring, normative constraints, LLM-mediated interaction, and an exploratory empirical evaluation with military professionals. Part of this framework is DIALOG, an LLM-mediated interaction component that translates natural language into doctrine-aligned plays, and builds up a context with ethical, legal, and operational constraints. We evaluate this framework in TNO's vehicle simulation facility, where five military professionals executed tactical missions using closed play, open play, and goal-based delegation modes. Findings reveal a nuanced autonomy-control trade-off: operators initially favored higher-autonomy modes but reverted to more structured control when confronted with unexpected or non-doctrinal agent behaviors. The results highlight implications for MHC, transparency, and trust calibration, and point toward research directions for enhancing tactical competence, communication fidelity, and verifiability in GPAI-enabled human-machine teams.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70091","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149146049","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":"Toward an enhanced scalability framework for human-multi-agent teams","authors":"Lawrence Dale Perkins,&nbsp;Hakki Erhan Sevil,&nbsp;Matthew Johnson","doi":"10.1002/aaai.70099","DOIUrl":"https://doi.org/10.1002/aaai.70099","url":null,"abstract":"<p>As autonomous systems become increasingly capable, the deployment of Human-Multi-Agent Teams (HMATs) offers immense potential for executing complex, decentralized tasks. However, realizing this potential is constrained by a critical bottleneck: the cognitive and supervisory capacity of the human operator. While foundational <i>fan-out</i> models established crucial groundwork for predicting this capacity, they are structurally limited to sequential, one-to-one interactions and rely on ambiguous variables that often produce overly optimistic predictions in real-world environments. This article synthesizes a multi-year empirical research program to present a comprehensive, new framework for modeling human-autonomy scalability. First, we deconstruct the variables of legacy models into human and robot elements, demonstrating that true scalability is a relational property driven by human-robot alignment rather than raw machine autonomy. We then validate the necessity of a mathematically bounded architecture to ensure predictive stability under operational uncertainty. Finally, we extend this framework to mathematically capture the paradigm shift toward group-based, one-to-many teaming. By analyzing the results of a complex, multi-agent simulation, we reveal the fundamental, quantifiable trade-off between increasing team scale and maintaining human oversight, providing AI system designers with a validated blueprint for optimizing the next generation of human-autonomy teams.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70099","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149146048","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":"PRECISE: Reducing the bias of LLM evaluations using prediction-powered ranking estimation","authors":"Abhishek Divekar,&nbsp;Anirban Majumder","doi":"10.1002/aaai.70066","DOIUrl":"https://doi.org/10.1002/aaai.70066","url":null,"abstract":"<p>Evaluating the quality of search systems traditionally requires a significant number of human relevance annotations. In recent times, several systems have explored the usage of Large Language Models (LLMs) as automated judges for this task while their inherent biases prevent direct use for metric estimation. We present a statistical framework extending Prediction-Powered Inference (PPI) (Angelopoulos, Duchi, and Zrnic 2024) that combines minimal human annotations with LLM judgments to produce reliable estimates of metrics which require sub-instance annotations. Our method requires as few as 100 human-annotated queries and 10,000 unlabeled examples, reducing annotation requirements by significantly compared to traditional approaches. We formulate our proposed framework (<span>PRECISE</span>) for inference of relevance uplift for an LLM-based query reformulation application, extending PPI to sub-instance annotations at the query-document level. By reformulating the metric-integration space, we reduced the computational complexity from <span></span><math>\u0000 <semantics>\u0000 <mrow>\u0000 <mi>O</mi>\u0000 <mo>(</mo>\u0000 <msup>\u0000 <mn>2</mn>\u0000 <mrow>\u0000 <mo>|</mo>\u0000 <mi>C</mi>\u0000 <mo>|</mo>\u0000 </mrow>\u0000 </msup>\u0000 <mo>)</mo>\u0000 </mrow>\u0000 <annotation>$O(2^{|C|})$</annotation>\u0000 </semantics></math> to <span></span><math>\u0000 <semantics>\u0000 <mrow>\u0000 <mi>O</mi>\u0000 <mo>(</mo>\u0000 <msup>\u0000 <mn>2</mn>\u0000 <mi>K</mi>\u0000 </msup>\u0000 <mo>)</mo>\u0000 </mrow>\u0000 <annotation>$O(2^K)$</annotation>\u0000 </semantics></math>, where <span></span><math>\u0000 <semantics>\u0000 <mrow>\u0000 <mo>|</mo>\u0000 <mi>C</mi>\u0000 <mo>|</mo>\u0000 </mrow>\u0000 <annotation>$|C|$</annotation>\u0000 </semantics></math> represents corpus size (in order of millions). Detailed experiments across prominent retrieval datasets demonstrate that our method reduces the variance of estimates for the business-critical Precision@<i>K</i> metric, while effectively correcting for LLM bias in low-resource settings.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70066","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149146090","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":"Structural limits of decision attribution in modular automated governance systems","authors":"Manuele Tele Junior Fernandez","doi":"10.1002/aaai.70101","DOIUrl":"https://doi.org/10.1002/aaai.70101","url":null,"abstract":"<p>Modular automated governance systems increasingly coordinate distributed decision-making across multiple agents. In blockchain-based DAOs and automated on-chain deployment pipelines, all state transitions are publicly observable. This transparency is frequently equated with accountability. We challenge this assumption. We argue that in certain modular governance architectures, full observability of execution does not guarantee univocal decision attribution. This limitation arises not from missing data or intentional privacy mechanisms, but from structural properties that separate deliberation, delegation, and execution without encoding deliberative provenance into the observable state. We frame this as a conditional non-injectivity of the mapping between deliberative configurations and observable traces. Through case studies of real DAO governance executions, automated deployment pipelines, and controlled experiments on LLM-based multi-agent systems, we provide empirical evidence that distinct deliberative configurations can produce indistinguishable observable outcomes. We do not claim that attribution is universally impossible, but identify a structural regime in which transparency does not guarantee recoverability of deliberative control. We discuss the implications for accountability and safety in multi-agent systems.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70101","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149146062","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":"Designing healthy and resilient information environments: A multi-agent sandbox for exploring risks and countermeasures","authors":"Jurriaan van Diggelen,&nbsp;Maaike Homan","doi":"10.1002/aaai.70098","DOIUrl":"https://doi.org/10.1002/aaai.70098","url":null,"abstract":"<p>Artificial intelligence is transforming online information environments, amplifying both societal benefits and risks. This article argues that online platforms can be understood as multi-agent systems (MAS) composed of humans, simulated users, adversarial agents, and defensive agents operating under platform-level rules. In this MAS-based framework, persuasion, trust, coordination, and governance interact to shape system-level outcomes. To study these dynamics safely, we introduce a sandbox environment in which human participants, red bots, green bots, and blue bots can be observed under controlled conditions. Early experiences with the sandbox show that persuasive agents can be built with little effort, that current large language model (LLM) agents lack psychological realism, and that humans may struggle to distinguish malicious red bots from benign participants. The sandbox enables stakeholders to experience and evaluate trade-offs in moderation, amplification, and intervention strategies. We outline a research agenda for improving agent validity, modelling advanced threats, and designing human–machine teams that support healthier, more resilient information environments.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70098","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149145668","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}