Association for theAdvancement ofArtificial IntelligenceAAAI 2025 PRESIDENTIAL PANEL ON THEFuture of Al ResearchPublished March 2025Association for theAdvancement ofArtificial IntelligenceAAAI 2025 PRESIDENTIAL PANEL ON THEFuture of Al ResearchPublished March 2025:Table of Contents7Introduction10Panel Members Contributors12Al Reasoning16Al Factuality Trustworthiness20Al AgentsAl Evaluation28Al Ethics Safety33Embodied Al37Al Cognitive Science41Hardware Al45Al for Social Good49Al Sustainability56Al for Scientific Discovery61 Artificial General Intelligence(AGI)67Al Perception vs.Reality71 Diversity of Al Research Approaches75Research Beyond the Al Research Community79Role of Academia83Geopolitical Aspects Implications of Al位IntroductionAs Al capabilities evolve rapidly,Al research is alsoresilience of the peer-review system,with theundergoing a fast and significant transformationimmediate release of papers without peer-reviewalong many dimensions,including its topics,evaluation having become widely accepted acrossits methods,the research community,and themany areas of Al research.Legacy and social mediaworking environment.Topics such as Al reasoningincreasingly cover Al research advancements.and agentic Al have been studied for decades butoften with contradictory statements that confusenow have an expanded scope in light of current Althe readers and blur the line between reality andcapabilities and limitations.Al ethics and safety,Alperception of Al capabilities.All this is happeningfor social good,and sustainable Al have becomein a geo-political environment,in whichcentral themes in all major Al conferences.companies and countries compete fiercely andMoreover,research on Al algorithms andglobally to lead the Al race.This rivalry may impactsoftware systems is becoming increasingly tiedaccess to research results and infrastructure asto substantial amounts of dedicated Al hardwarewell as global governance efforts,underscoring thenotably GPUs,which leads to Al architecture co-need for international cooperation in Al researchcreation,in a way that is more prominent nowand innovation.than over the last 3 decades.Related to this shift.more and more Al researchers work in corporateIn this overwhelming multi-dimensional andenvironments,where the necessary hardwarevery dynamic scenario,it is important to be ableand other resources are more easily available,to clearly identify the trajectory of Al research incompared to academia,questioning the rolesa structured way.Such an effort can define theof academic Al research,student retention,andcurrent trends and the research challenges stillfaculty recruitingahead of us to make Al more capable and reliable,so we can safely use it in mundane but also,mostThe pervasive use of Al in our daily lives and itsimportantly,in high-stake scenarios.impact on people,society,and the environmentmakes Al a socio-technical field of study,thusThis study aims to do this by including 17 topicshighlighting the need for Al researchers to workrelated to Al research,covering most of thewith experts from other disciplines,such astransformations mentioned above.Each chapterpsychologists,sociologists,philosophers,andof the study is devoted to one of these topics,economists.The growing focus on emergent Alsketching its history,current trends and openbehaviors rather than on designed and validatedchallengesproperties of Al systems renders principledTo conduct this study,I selected a very diverseempirical evaluation more important thangroup of 24 experienced Al researchers,whoever.Hence the need arises for well-designedgenerously accepted my invitation and devotedbenchmarks,test methodologies,and sounda significant amount of time to this effort.Weprocesses to infer conclusions from the results ofall worked together between summer 2024 andcomputational experiments.The exponentiallyspring 2025 to structure the study,define theincreasing quantity of Al research publicationsmain topics,discuss the content,comment andand the speed of Al innovation are testing thecontribute to the various chapters.Additionally,some chapters engaged also withThe work around the entire study has beenadditional contributors who brought theirgenerously supported and made possible by theexpertise on a specific topic.The work was doneamazing work of Meredith Ellison,AAAl Executivemostly online,with monthly calls with all panelDirector,and the AAAl office staff,who alsomembers plus additional calls for the teamprepared and delivered the survey.working on each chapter,with also in a full-day in-person meeting.held in January 2025I hope that this report will be useful to the wholeAl research community.However,the report hasHowever,we also wanted to include the opinionbeen intentionally written in a non-technicalof the entire AAAl community,so we launchedway,to reach out to other audiences,includingan extensive survey on the topics of the study,experts of other disciplines,policy makers,fundingwhich engaged 475 respondents,of which aboutagencies,the media,and the general public.20%were students.Among the respondents,We all need to work together to advance Al in aacademia was given as the main affiliation(67%).responsible way,to make sure that technologicalfollowed by corporate research environmentprogress supports the progress of humanity and is(19%).Geographically,the most represented areasaligned to human valuesare North America(53%).Asia(20%),and Europe(19%).While the vast majority of the respondentslisted Al as one of their primary fields of study.there were also mentions of other fields,suchas neuroscience,medicine,biology,sociology,Francesca Rossiphilosophy,political science,and economics.ThisAAAl President,2022-2025interest in multi-disciplinary research from 95%ofthe respondents.Each chapter of this report includes a briefsummary of the responses to questions related tothe respective topic.The panel's findings are opinions of the panel members and do not represent theopinion of their institutions or companies.运营动脉运营动脉运营动脉运营动脉运营动脉运营动脉运营动脉运营动脉运营动运营动脉运营动脉运营动脉运营动脉运营动脉运营动脉运营动脉运营动脉运营动脉运营动脉运营动脉运营动脉运营动脉运营动脉运营动脉运营动运营动脉运营动脉运营动运营动9Panel Members AdditionalContributorsPanel MembersFrancesca Rossi,Eugene Freuder,Alan Mackworth,IBM ResearchUniversity College CorkUniversity of BritishColumbiaChristian Bessiere,Yolanda Gil,University of MontpellierUniversity of SouthernKaren Myers,CaliforniaSRI InternationalJoydeep Biswas,University of Texas at AustinHolger Hoos,Luc De RaedtRWTH Aachen University,KU Leuven and OrebroRodney BrooksGermany and LeidenUniversityMassachusetts Institute ofUniversity,The NetherlandsTechnologyStuart Russell,Eric Horvitz.University of CalifomiaVincent Conitzer,MicrosoftBerkeleyCarnegie Mellon UniversitySubbarao Kambhampati,BartSelman,Thomas G.Dietterich,Arizona State UniversityCornell UniversityOregon State UniversityHenry KautzPeter StoneVirginia Dignum,University of VirginiaThe University of Texas atUmea UniversityAustin and Sony AlJihie Kim,Oren Etzioni,Dongguk UniversityMillind Tambe,University of WashingtonHarvard UniversityHiroaki Kitano,Kenneth D.Forbus,Sony ResearchMichael Wooldridge,Northwestern UniversityUniversity of OxfordAdditional ContributorsAditya Akella,Norm Jouppi,Yoav Shoham,University of Texas at AustinGoogleStanford UniversityChapter:Hardware AlChapter:Hardware and AlChapter:Al AgentsYoshua Bengio,John E.Laird,Carles Sierra,MILAUniversity of MichiganSpanish National ResearchChapter:Artificial GeneralChapter:Al CognitiveCouncilScienceChapter:Al AgentsAbeba Birhane,Amy Luers,Pradeep Varakantham,Trinity College DublinMicrosoftSingapore ManagementChapter:Research BeyondChapter:Al SustainabilityUniversitythe Al Research CommunityChapter:Al for Social GoodPeter Norvig.Bill Dally,GoogleNVIDIAChapter:Hardware and AlIntelligence (AGI)Besmira Nushi,Carnegie Mellon UniversityMicrosoft ResearchChapter:Al for Social GoodIntelligence (AGI)Jonathan Gratch,University of SouthernBalaraman RavindranCaliforniaIndian Institute ofChapter:Al CognitiveTechnology MadrasScienceChapter:Al for Social GoodAl ReasoningCHAIRSChristian Bessiere,University of MontpellierThe ability to reason has been a salient characteristic of humanHolger Hoos,RWTH Aachen University,intelligence,and there is a critical need for verifiable reasoningGermany and Leiden Univer-in Al systems.sity,The NetherlandsSubbarao Kambhampati,Arizona State UniversityMain TakeawaysReasoning has always been seen as a core characteristic of human intelligenceReasoning is used to derive new information from given base knowledge;thisnew information is guaranteed correct when sound formal reasoning is used,otherwise it is merely plausible.Al research has led to a range of automated reasoning techniques.Thesereasoning techniques have given rise to Al algorithms and systems,includingSAT,SMT,and constraints solvers as well as probabilistic graphical models,all ofwhich play a key role in critical real-world applications.While large pre-trained systems(such as LLMs)have made impressiveadvancements in their reasoning capabilities,more research is needed toguarantee correctness and depth of the reasoning performed by them;suchguarantees are particularly important for autonomously operating Al agents12Al ReasoningContext Historyresearch have covered the gamut frompatterns as they emerge automaticallyplanning and temporal reasoning toafter large-scale training on petabyteReasoning is a core component ofdiagnosis and explanation.While earlycorpora.While the results have beenhuman intelligence.From the dawnAl has paid attention to both plausiblequite remarkable so far,the reasoning inof humanity,abductive reasoningreasoning (case-based,analogical,this context has been of the "plausible"has been used to predict danger andqualitative)and sound formal reasoningvariety with no guarantees.inductive reasoning made it possiblewith guarantees(logical,probabilistic.to learn regularities governing theconstraint-based),over the years,Meanwhile,sound formal reasoningworld.Beginning in Ancient Greece,the focus has shifted more towardstechniques remain key to importantdeductive reasoning techniques werereasoning with formal guarantees.and impactful applications of cutting-developed to draw valid conclusionsThere are good reasons for this whenedge Al technology for the verificationdesigning Al systems and techniquesof computer hardware and software,known to be true.The developmentthat compensate for human limitationsas well as for real-world planning andof reasoning methods with such aand weaknesses since reasoningresource allocation problems.Theypriori guarantees was a key factor inwith guarantees is challenging forare also increasingly recognized as athe advancement of modern science,humans.This has led to practicallycrucial basis for the formal verificationmathematics,and engineering;notably,impactful applications of Al systemsof machine learning techniques suchaccording to philosophers such assuch as SAT,SMT,and constraintsas neural networks,e.g.,in the contextCharles Sanders Peirce,the interplaysolvers,including the verification ofof local robustness against adversarialbetween abduction,deduction,andcorrectness properties of computerattacks [6].Significant research activityinduction forms the basis of thehardware and software,the safety oftakes place in these areas,focusing onscientific method and hence all moderncommunications protocols,the designimproving various types of reasoningscience.Attempts to mechanizeof new proteins,and,more recently,thealgorithms(notably with respect to theirlogical reasoning can be traced back torobustness of neural networks againstcomputational complexity),leveraging13th-century philosopher Ramon Lulladversarial attacks.It has also resultedlearning within sound formal reasoning.and lie at the heart of the concept ofin probabilistic graphical models [4.and combining reasoning and learningcomputation.Probabilistic reasoning5],which are powerful modeling andtechniques [7,8].and inference have also profoundlyinference tools that have found theirimpacted reasoning,often relying onway into numerous applications ofthe celebrated theorem by Thomasreasoning in medicine,robotics,andResearch ChallengesBayes on inverse probability that alsobeyond.Bringing some of the rigorous a priori orforms the basis for many machinepost hoc guarantees back into plausiblelearning and statistics approaches.Current State Trendsreasoning patterns turbocharged byFinally,the evaluation of correct(sound)the pre-trained models has become anreasoning lies at the heart of mostactive and promising area of researchquantitative assessments of humanThe emergence of the Internet and theassociated technology that made itespecially where Al systems need tocognition.possible to capture the human digitalwork autonomously in safety-criticalNot surprisingly,reasoning has beenfootprint at scale,as well as the leaps indomains.Research on so-called "largereasoning models"as well as on neuro-central to the Al enterprise.Indeed,computing power,have made possiblethe earliest research in Al-from Logicnovel approaches to learning bottom-symbolic approaches is addressingTheorist onwards [1]-had a strongup from data.Of particular interest arethese challenges.focus on reasoning [2].Since the 1960s,large pre-trained models,such as LLMs,Furthermore,even though formalAl has also embraced probabilisticthat have shown surprising abilities inreasoning with correctness guaranteesreasoning and models,initially forplausible reasoning.Unlike the earlieris currently considerably less in voguemedical diagnosis [3].Since then,research on reasoning in Al,LLMsthan the use of generative Al techniquesthe reasoning tasks addressed in Alhave focused on plausible reasoningfor plausible reasoning,formidable andAl Reasoningessential challenges also remain in thatHow can computers betterarea.In this context,the combinationunderstand and simulate humanof machine learning techniques withreasoning?formal reasoning techniques holdsWhat is the role of collaborativeconsiderable promise for economicallyreasoning between humans andand socially valuable breakthroughs,computers?notably in the area of Al safety andtransparencyHow best can LLMs and symbolicreasoning be integrated into "neuro-The questions and challenges we facesymbolic reasoning'"？range from the philosophical:Are further breakthroughs,beyondWhat exactly is"reasoning"?both LLMs and traditional symbolicreasoning,required to achieve AGl-to the practical:level reasoning?Can LLM'reasoning'be trusted?What forms of reasoning can bestand include:support humans when dealing withvarious challenges,e.g.,in medical,What does the future hold for thescientific,engineering,and legaladvancement and role of symbolicdomains?reasoning?To what extent can LLMs or othergenerative models reproduce orreplace symbolic reasoning?To what degree will symbolicreasoning be necessary or sufficientto overcome the current limitationsof LLMs？How well can Al reasoning,especially LLM'reasoning,'beexplained and understood?1.Newell.A.Simon.H.(1956)The logic theory machine:Acomplex information processing system.IRE Transactionson Infor mation Theory 2:61-792.Brachman.R.and Levesque.H.(2004)Knowledge Representation and Reasoning (1st Ed).Morgan Kaufman6.Konig.M et al.(2024)Critically Assessing the State of the Art in Neural Network Verification.Journal of Machine Learning Research 25(12):1-537.Guo.D.et.al.(2025)DeepSeek-R:Incentivizing Reasoning Capability in LLMs via Reinforcement Learning-https//arxiv.org/abs/2501129488.Kambhampati.S.(2024)Can Large Language Models Reason and Plan?Annals of New York Academy of Sciences.March 2024.Al ReasoningCommunity OpinionThe AAAl community appears towarranted to better communicate theof learning and reasoning approachesimportance and success of formal,as very important(6 or 7 on a scaleof reasoning in Al systems.In oursound reasoning techniques.Finally,of 7);interestingly,the percentagecommunity survey,slightly over 55%44.7%of respondents agreed thatof respondents that consideredof the respondents chose to answer"Reasoning involves a search process."Explainability and verifiability as veryspecific questions related to the topicimportant was similarly high(at 71.7%).of reasoning.Of these,79%indicatedThere was broad agreement amongthat the topic of reasoning is relevantsurvey participants that focusingFinally,61.8%of survey participantsto their research(with 44.7%markingreasoning research in Al on human-estimated the minimal percentage ofit as"very relevant").Of the propertieslevel reasoning is valuable(41.6%)orsymbolic Al techniques required forrequired for referring to a processeven essential(47%);similarly,a focusreaching human-level reasoning to beas reasoning,77.5%of the surveyon domain-specific reasoning abilitiesat least 50%(with 24.8%estimating itparticipants marked"Knowledge canwas seen by 49.6%of respondents asat 75%or more,compared to 38.2%be incorporated",72.5%"Explanationsvaluable,and by 42.8%as essentialestimating it at 25%or below).Whatcan be provided,"and 56.9%"InvolvesThis clearly reflects the importanceremains unclear is the degree tomultiple steps to arrive at a conclusion".attributed to a research focus onwhich Al researchers and practitionersInterestingly,merely 37.4%indicatedreasoning.realize that decidedly superhuman'Guaranteed correctness of inferencelevels of reasoning are required forThe community also sees an excitingresults/outcomes",and only 23.7%thatand displayed in the prominent andpotential of synergy offered by logical"A formal system and solver is used,"successful applications of formal Aland probabilistic models of reasoningwhich reflects the recent focus onthat were developed in Al prior to largeinformal,plausible reasoning,likely inand mathematical discovery andpre-trained models.This is clearlythe context of generative Al methods.engineering applications,as well as in Alreflected in the fact that 76.9%of surveyThis suggests that an effort may besafety.participants marked the integration15