2025关于人工智能研究未来研究报告

A A AI 2025 PRESIDENTIAL PANEL ON THE Future of AI Research Published March 2025 Table of Contents 7Introduction10Panel Members & Contributors12AI Reasoning16AI Factuality & Trustworthiness20AI Agents24AI Evaluation28AI Ethics & Safety33Embodied AI37AI & Cognitive Science41Hardware & AI45AI for Social Good49AI & Sustainability56AI for Scientific Discovery61Artificial General Intelligence (AGI)67AI Perception vs. Reality71Diversity of AI Research Approaches75Research Beyond the AI Research Community79Role of Academia83Geopolitical Aspects & Implications ofAI Introduction resilience of the peer-review system, with theimmediate release of papers without peer-reviewevaluation having become widely accepted acrossmany areas of AI research. Legacy and social mediaincreasingly cover AI research advancements,often with contradictory statements that confusethe readers and blur the line between reality andperception of AI capabilities. All this is happeningin a geo-political environment, in whichcompanies and countries compete fiercely andglobally to lead the AI race. This rivalry may impactaccess to research results and infrastructure aswell as global governance efforts, underscoring theneed for international cooperation in AI researchand innovation. As AI capabilities evolve rapidly, AI research is alsoundergoing a fast and significant transformationalong many dimensions, including its topics,its methods, the research community, and theworking environment. Topics such as AI reasoningand agentic AI have been studied for decades butnow have an expanded scope in light of current AIcapabilities and limitations. AI ethics and safety, AIfor social good, and sustainable AI have becomecentral themes in all major AI conferences.Moreover, research on AI algorithms andsoftware systems is becoming increasingly tiedto substantial amounts of dedicated AI hardware,notably GPUs, which leads to AI architecture co-creation, in a way that is more prominent nowthan over the last 3 decades. Related to this shift,more and more AI researchers work in corporateenvironments, where the necessary hardwareand other resources are more easily available,compared to academia, questioning the rolesof academic AI research, student retention, andfaculty recruiting. In this overwhelming multi-dimensional andvery dynamic scenario, it is important to be ableto clearly identify the trajectory of AI research ina structured way. Such an effort can define thecurrent trends and the research challenges stillahead of us to make AI more capable and reliable,so we can safely use it in mundane but also, mostimportantly, in high-stake scenarios. The pervasive use of AI in our daily lives and itsimpact on people, society, and the environmentmakes AI a socio-technical field of study, thushighlighting the need for AI researchers to workwith experts from other disciplines, such aspsychologists, sociologists, philosophers, andeconomists. The growing focus on emergent AIbehaviors rather than on designed and validatedproperties of AI systems renders principledempirical evaluation more important thanever. Hence the need arises for well-designedbenchmarks, test methodologies, and soundprocesses to infer conclusions from the results ofcomputational experiments. The exponentiallyincreasing quantity of AI research publicationsand the speed of AI innovation are testing the This study aims to do this by including 17 topicsrelated to AI research, covering most of thetransformations mentioned above. Each chapterof the study is devoted to one of these topics,sketching its history, current trends and openchallenges. To conduct this study, I selected a very diversegroup of 24 experienced AI researchers, whogenerously accepted my invitation and devoteda significant amount of time to this effort. Weall worked together between summer 2024 andspring 2025 to structure the study, define themain topics, discuss the content, comment andcontribute to the various chapters. The work around the entire study has beengenerously supported and made possible by theamazing work of Meredith Ellison, AAAI ExecutiveDirector, and the AAAI office staff, who alsoprepared and delivered the survey. Additionally, some chapters engaged also withadditional contributors who brought theirexpertise on a specific topic. The work was donemostly online, with monthly calls with all panelmembers plus additional calls for the teamworking on each chapter, with also in a full-day in-person meeting, held in January 2025. I hope that this report will be useful to the wholeAI research community. However, the report hasbeen intentionally written in a non-technicalway, to reach out to other audiences, includingexperts of other disciplines, policy makers, fundingagencies, the media, and the general public.We all need to work together to advance AI in aresponsible way, to make sure that technologicalprogress supports the progress of humanity and isaligned to human values. However, we also wanted to include the opi