SCOPE: What Counts as Discovery in the Age of AI?

About the workshop

When an AI system produces a new mathematical proof, proposes an experimentally validated hypothesis, or carries out a computational research project, what should count as its contribution to science?

The answer affects more than authorship. It shapes what researchers investigate, what reviewers accept as evidence, how institutions allocate credit, and whether scientific communities can understand and build on the resulting work. As AI takes on a larger role in research, these questions require discussion among the people developing the systems and the communities whose practices they may change.

SCOPE will bring together researchers developing AI systems, scientists using and evaluating their outputs, and scholars who study how science is done. Together, we will examine what AI contributes to research, what evidence supports its results, what humans need to understand, how credit should be allocated, and how findings should be published. We will begin with mathematics and draw on empirical science and machine learning to explore which principles can be shared across disciplines and where differences in research practices call for different standards.

Why this discussion now?

Recent developments in mathematics make this discussion immediate. On October 6, 2026, OpenAI released a collection of mathematics manuscripts generated by an internal model, along with details of the research process and formalizations of selected proofs (OpenAI, 2026). The release exposed disagreement over what constitutes mathematical progress. In a statement republished on Terence Tao's blog, the Association for Human Mathematics rejected the claim that this publication model advances mathematics, describing the mass release as a “demonstration of power” and calling for science that “centers human understanding” (Association for Human Mathematics Communications Working Group, 2026). The independent Advisory Group on Mathematics and Artificial Intelligence likewise distinguishes the release of mathematical results from the work needed to understand and integrate them into established knowledge, while continuing to engage with AI labs (Advisory Group on Mathematics and Artificial Intelligence, 2026b).

At the heart of this disagreement is the question: how should scientific communities assess a result's validity, its contribution to understanding, and the conditions under which it was produced and released?

The same questions extend to empirical science and machine learning. AI research systems now carry out idea generation, computational experimentation, and manuscript preparation (Lu et al., 2026), while AI co-scientist systems have been developed to propose biomedical hypotheses for experimental investigation (Gottweis et al., 2025). These efforts pursue a longstanding ambition to develop AI capable of scientific discovery (Kitano, 2021). They also make it essential to distinguish evidence for a scientific finding from evidence for the capabilities of the system involved. A useful discovery may depend on substantial human judgment, and a technically impressive system may produce results of limited scientific significance.

Researchers differ in how they envision this transition. Tao (2026) examines mathematical values and practices in the age of AI, while Szegedy (2026) anticipates changes in the kinds of mathematical work that receive recognition. Jimenez et al. (2026) argue for developing AI scientists alongside institutional reforms that support verification and accountability. Messeri & Crockett (2024) warn that greater productivity can coexist with illusions of understanding and a narrowing of scientific inquiry.

The debate concerns both what counts as scientific progress and how researchers participate in it: whether valid new results are sufficient, what role human understanding should play, and whether scientific communities can collaborate with frontier AI labs while retaining their independence. SCOPE will examine these questions without assuming that more automation is necessarily desirable, or that current research practices should remain unchanged.

Invited speakers and panelists

The participants below connect the workshop's questions to mathematical practice, empirical science, AI evaluation, and the development of autonomous research systems.

Pushmeet Kohli

Vice President of Research

Google DeepMind Confirmed

Pushmeet Kohli is Vice President of Research at Google DeepMind and Chief Scientist of Google Cloud. He founded and leads DeepMind's AI for Science and Strategic Initiatives Unit, whose work includes AlphaFold for protein structure prediction and AlphaProof for formal mathematical reasoning. He coauthored AlphaEvolve: A Coding Agent for Scientific and Algorithmic Discovery, which uses language models and automated evaluators to discover and improve algorithms, and Accelerating Scientific Discovery with Co-Scientist, which investigates multi-agent hypothesis generation with experimental validation in biomedical research. Previously, he was Director of Research for the Cognition group at Microsoft Research. His experience leading the development of these systems would bring an industry perspective on what evidence is needed to move from a model-generated candidate to a scientific result, how domain experts shape that process, and how laboratories should communicate novelty, limitations, and human involvement.

Tengyu Ma

Assistant Professor of Computer Science

Stanford University Confirmed

Tengyu Ma is an Assistant Professor of Computer Science at Stanford University who works on learning theory, foundation models, reinforcement learning, robustness, and non-convex optimization. His distinctions include a Sloan Research Fellowship, an NSF CAREER Award, and a COLT Best Paper Award. He coauthored Priced Guidance: Can Language Models Generate Future Research Ideas? (Wen et al., 2026), which connects theoretical analysis with empirical evaluation of research-idea generation. The framework relates the information supplied to a model to a lower bound on its probability of generating a target idea without guidance. His perspective would help examine what formal guarantees and empirical measurements can establish about novelty, reasoning, and generalization, and where the assumptions of an evaluation leave questions of scientific originality unresolved.

François Charles

Professor of Mathematics

École normale supérieure–PSL Confirmed

François Charles is a Professor of Mathematics at École normale supérieure–PSL and editor-in-chief of the Journal de l'École polytechnique—Mathématiques. His research in arithmetic and algebraic geometry includes K3 surfaces, Hodge theory, and the Tate conjecture, with work such as The Tate Conjecture for K3 Surfaces over Finite Fields published in Inventiones Mathematicae. He is a member of the independent Advisory Group on Mathematics and Artificial Intelligence, whose Responsible Release of AI-Generated Mathematics recommendations (Advisory Group on Mathematics and Artificial Intelligence, 2026a) address how AI laboratories should engage with mathematical research. The group's October 6 statement (Advisory Group on Mathematics and Artificial Intelligence, 2026b) distinguishes publication from the subsequent work of understanding and incorporating results into mathematical knowledge. His research and editorial experience will ground discussion of proof assessment, scholarly credit, editorial responsibility, and the standards journals should apply to AI-assisted mathematics.

Licia Verde

Professor of Cosmology

ICREA / Institute of Cosmos Sciences (ICCUB-IEEC), Facultat de Física, University of Barcelona Confirmed

Licia Verde is an ICREA Research Professor at the University of Barcelona's Institute of Cosmos Sciences. Her research covers theoretical cosmology, the cosmic microwave background, large-scale structure, galaxy surveys, and statistical inference. Her research distinctions include the 2012 Gruber Cosmology Prize and the 2018 Breakthrough Prize in Fundamental Physics, and she serves as editor and scientific director of the Journal of Cosmology and Astroparticle Physics. She coauthored AI Scientists as Engines of Discovery: A Case for Development within Reformed Institutions (Jimenez et al., 2026), which connects the development of autonomous research systems to reforms in verification, accountability, interpretability, authorship, and peer review. Her perspective would connect scientific autonomy to the practical standards of empirical science and the continuing role of domain expertise, including how scientific institutions should assess and incorporate discoveries made with AI.

Sayash Kapoor

Incoming Assistant Professor, School of Information

UC Berkeley Confirmed

Sayash Kapoor is an incoming Assistant Professor at the UC Berkeley School of Information, starting in July 2027. He studies AI evaluation, reproducibility, and the relationship between technical capabilities and their use in scientific institutions. He received his PhD in Computer Science at Princeton University in August 2026. He coauthored AI Snake Oil with Arvind Narayanan and Leakage and the Reproducibility Crisis in Machine-Learning-Based Science (Kapoor & Narayanan, 2023). His work includes CORE-Bench, which evaluates agents on reproducing computational research, and the Science of Agent Evaluation research program. These contributions connect critical examination of scientific AI claims with concrete ways to measure useful research capabilities. He would help distinguish benchmark success, reproducible computation, and evidence of scientific capability, and examine what evaluation results actually establish about an agent's contribution.

Marinka Zitnik

Associate Professor of Biomedical Informatics

Harvard Medical School Confirmed

Marinka Zitnik is an Associate Professor of Biomedical Informatics at Harvard Medical School, Associate Faculty at Harvard's Kempner Institute, and an Associate Member of the Broad Institute of MIT and Harvard. She develops AI methods for therapeutic discovery, genomic medicine, and reasoning over biomedical knowledge, with an emphasis on systems that contribute to scientific understanding. She founded Therapeutics Data Commons and leads the international AI4Science initiative. Her work includes TxGNN: A Foundation Model for Clinician-Centered Drug Repurposing, which combines graph learning with tools that help clinicians interpret candidate therapeutic predictions. This work provides a concrete setting for asking how generated hypotheses relate to biological evidence, expert interpretation, and further experimental validation. She would connect the workshop's discussion of novelty and contribution to useful biomedical findings and the evidence required to support them.

Hiroaki Kitano

Chief Technology Fellow; President and CEO

Sony Group / Sony Computer Science Laboratories Confirmed

Hiroaki Kitano is Chief Technology Fellow at Sony Group Corporation and President and CEO of Sony Computer Science Laboratories. His research brings together artificial intelligence, robotics, and systems biology. He helped establish systems biology as a field that examines how biological components interact and function together. He received the IJCAI Computers and Thought Award, is a Fellow of the Association for the Advancement of Artificial Intelligence, and founded the RoboCup Federation. His Nobel Turing Challenge: Creating the Engine for Scientific Discovery (Kitano, 2021) sets out a long-term program for highly autonomous systems capable of scientific work comparable to that of leading human researchers. He would bring an industry research perspective on scientific autonomy, ambitious discovery goals, and the standards by which such systems should be judged, informed by experience building research programs across AI and the life sciences.

Sara Beery

Assistant Professor, Artificial Intelligence and Decision Making

Massachusetts Institute of Technology Confirmed

Sara Beery is an Assistant Professor in MIT's Faculty of Artificial Intelligence and Decision Making and CSAIL, where she develops computer vision methods for environmental and biodiversity monitoring. Her research addresses geographic and temporal distribution shifts, imperfect observations, and the difficulty of turning model predictions into reliable ecological knowledge. She received an NSF CAREER Award and a Schmidt Sciences AI2050 Early Career Fellowship, founded the AI for Conservation community, and co-leads the NSF/NSERC Global Center on AI and Biodiversity Change. Her essay AI & Ecology: From Tool to Transformation (Beery, 2026) and coauthored article Generative AI as a Tool to Accelerate the Field of Ecology examine how AI changes ecological research. She would bring a perspective grounded in field observation, domain knowledge, uncertainty, and whose expertise is recognized when AI contributes to a discovery.

Discussion questions

The workshop will address six questions, drawing on different perspectives across research communities.

What makes an AI-produced result a scientific contribution?

Are correctness and novelty sufficient for scientific progress, or must a contribution also deepen understanding, explain a phenomenon, or enable further inquiry? Can a problem be considered solved before researchers understand the ideas underlying its solution? How should we distinguish producing a valid result from making that result usable as shared scientific knowledge? Should the process by which a result is produced affect its scientific value, the credit it warrants, or both?

What evidence should support a claim of discovery?

What independent checks are needed before an AI-produced result is accepted as an advance? What should be disclosed about prior work, human intervention, unsuccessful attempts, and computational resources? Research on leakage in machine-learning-based science shows how apparently strong results can rest on flawed evaluation (Kapoor & Narayanan, 2023). How should standards of evidence differ for a scientific claim and a claim about the AI system that produced it? Meiburg (2026a) reports that some Comparator challenge specifications allow changes to definitions that should be fixed. How can evaluators establish both that a proof is valid and that it proves the intended statement under the intended assumptions?

What must humans understand and take responsibility for?

When is verification sufficient, and when is an explanation essential? Must an author understand every step, the main ideas, or the grounds for trusting contributions supplied by others? The mathematics release recommendations emphasize the distinction between verified output and human understanding (Advisory Group on Mathematics and Artificial Intelligence, 2026a). Who should answer questions, investigate errors, and provide corrections when understanding is distributed across people and systems?

How should authorship and credit be allocated?

Should AI systems appear as authors, be cited as tools, or be identified through contribution statements? How should records distinguish problem selection, generation, validation, formalization, interpretation, and exposition? What credit is due to prior research and data, or to people who check and explain a result after its release? How should attribution relate to accountability?

How should AI-generated work enter the scientific record?

Should existing venues revise their disclosure and review practices, or should dedicated repositories or publication tracks serve particular kinds of work? What should distinguish a deposited result, a verified claim, and a contribution that a community can understand and build upon? When AI systems produce results faster than experts can assess them, who should undertake, fund, and receive credit for verification, explanation, and correction (Advisory Group on Mathematics and Artificial Intelligence, 2026a)? Building on Leonelli (2023), how should publication practices preserve provenance and make the resulting knowledge usable? What responsibilities should accompany a large-scale release?

What should research evaluation reward?

If AI makes some outputs easier to produce, how should publication, hiring, funding, and model evaluation respond? How can they recognize problem formulation, careful validation, explanation, and work that enables subsequent discoveries? When measures such as theorem counts become optimization targets, how can evaluation still capture growth in understanding, the discovery of new questions, and the development of researchers' expertise (Meiburg, 2026b)? Ecology highlights the importance of field observation, domain expertise, and participation in AI-enabled research (Beery, 2026). Who should determine research priorities when frontier capabilities and computational resources are concentrated in a small number of laboratories? How can researchers retain the freedom to formulate their own questions, rather than primarily validating or explaining results selected by AI developers (Advisory Group on Mathematics and Artificial Intelligence, 2026b)? What would equitable participation require for early-career researchers and communities with limited access to data, compute, and AI tools?

Tentative one-day program

The program comprises two keynotes, one in the morning and one in the afternoon, four invited talks, four contributed talks, a dedicated 90-minute poster session, and a one-hour panel.

Time	Session	Proposed focus / format
09:00–09:15	Welcome and introduction	Workshop questions and agenda
09:15–09:55	Keynote: François Charles	When does a mathematical result become mathematical knowledge?
09:55–10:20	Invited talk: Tengyu Ma	How can we measure novelty and creativity in AI-generated research ideas?
10:20–10:35	Coffee break	☕ Informal conversation
11:00–11:30	Contributed talks I	Two papers: 10-minute talk + 5-minute Q&A each
11:30–12:15	Lunch break	🍽️ Lunch and informal exchange
12:15–13:45	Poster session	All accepted papers; discussion with authors
13:45–14:25	Keynote: Pushmeet Kohli	How do we establish that an AI system has made a scientific discovery?
14:25–14:50	Invited talk: Marinka Zitnik	When does an AI-generated hypothesis become a scientific finding?
14:50–15:05	Coffee break	☕ Informal conversation
15:05–15:30	Invited talk: Sara Beery	What can AI help us learn about the natural world?
15:30–16:00	Contributed talks II	Two papers: 10-minute talk + 5-minute Q&A each
16:00–16:15	Break	Informal conversation
16:15–17:15	Panel discussion	Who remains accountable when AI does the research? Moderated debate and audience questions
17:15–17:30	Closing remarks	Questions, disagreements, and possible next steps

The proposed panel brings together François Charles, Hiroaki Kitano, Licia Verde, and Sayash Kapoor, with a moderator from the organizing team, who will gather questions from the panelists beforehand to identify where their perspectives offer grounds for a healthy debate. Drawing on mathematical practice, research automation, AI evaluation, and empirical science, the discussion will explore where these perspectives agree and where they lead to different expectations of evidence, understanding, and responsibility. At least 15 minutes will be reserved for audience questions.

Organizing team

Our organizing team brings together researchers from mathematics, physics, and AI, whose complementary expertise, academic and industry perspectives, and experience organizing research events and building communities will shape both the scientific program and the running of the workshop.

Mariya I. Vasileva

Research Lead, Fundamental Research

Patronus AI

Mariya I. Vasileva leads the Fundamental Research team at Patronus AI and was until recently a Senior Research Scientist at Meta Superintelligence Labs, where she developed evaluation and oversight methods for frontier AI models. Her research focuses on scalable oversight and the capabilities of increasingly autonomous AI agents, alignment, multi-agent coordination and communication, and synthetic data generation. Her essay Recursive Self-Inflation examines how self-improvement loops can exploit a verifier's mistakes, and how to distinguish those gains from genuine progress. She holds a PhD in Computer Science from the University of Illinois Urbana-Champaign. Her organizing experience includes serving as General Chair of the Women in Machine Learning workshop at NeurIPS 2021 and organizing CVFAD at CVPR (2022–2024), ResistanceAI at NeurIPS 2020, and Broadening Participation in Data Mining at KDD 2017. She brings experience coordinating research programs and community participation, alongside an industry perspective on evaluating research agents.

Role: Lead organizer and designated ICLR contact, overseeing overall coordination, speaker relationships, sponsorship and funding, integration of the final program, and the workshop website.

Website: Personal homepage | Google Scholar: Profile

Jeremy Avigad

Dean's Chair in Logic and Philosophy of Mathematics

Carnegie Mellon University

Jeremy Avigad holds the Dean's Chair in Logic and Philosophy of Mathematics at Carnegie Mellon University, where he directs the Hoskinson Center for Formal Mathematics and the Institute for Computer-Aided Reasoning in Mathematics. His research bridges mathematical logic, formal verification, AI, and the philosophy of mathematical practice. He serves on the Lean FRO's board of directors and is an Amazon Scholar. His essay What Is Mathematics Now, and What Should It Be? argues for a broader conception of AI for mathematics and recognition of contributions to tools, libraries, and mathematical understanding. He also coauthored The Problem Is the Problem: Towards Scalable Mathematical Discovery, which addresses problem selection and expert review in human–AI research. His extensive organizing experience includes Machine Assisted Proof at IPAM and AI for Mathematics and Theoretical Computer Science at SLMath and the Simons Institute. He brings deep expertise in proof, explanation, and community standards to the organizing team.

Role: Scientific program advice and panel curation, refining discussion questions, ensuring a balance of perspectives, and helping shape the panel’s focus and participants.

Website: Personal homepage | Google Scholar: Profile

Publications and essays

Alex Meiburg

Assistant Professor of Applied Mathematics

University of Waterloo (IQC) / Perimeter Institute

Alex Meiburg is an Assistant Professor of Applied Mathematics at the University of Waterloo, affiliated with the Institute for Quantum Computing and the Perimeter Institute. His research connects quantum information, computational complexity, and formalized mathematics. He coauthored A Formalization of the Generalized Quantum Stein's Lemma in Lean, contributing to a foundation for computer-checked quantum theory. He also co-organized the 2026 AIM workshop on AI and number theory, bringing together academic mathematicians and industry researchers. In Thoughts on AI Solving “Math”, he argues that mathematical progress includes understanding, new questions, and the cultivation of expertise, beyond theorem production. His analysis of Comparator challenge specifications identifies a distinct concern: an accepted formal proof may establish a changed statement if its definitions are not protected. Together, this research, organizing experience, and public engagement connect the workshop's questions about scientific value to practical scrutiny of verification and evaluation.

Role: Program committee recruitment and mathematical review coordination, building an expert reviewer pool and matching contributions in mathematics and formal verification with appropriate reviewing expertise.

Website: Personal homepage | Google Scholar: Profile

Blog | Research updates

Claudio Mayrink Verdun

Research Scientist; Research Associate

Massachusetts Institute of Technology / Harvard University

Claudio Mayrink Verdun is a Research Scientist at MIT CSAIL, affiliated with MIT FutureTech and the Algorithmic Alignment Group, and a Research Associate at Harvard University. He holds a doctorate from the Technical University of Munich and works across mathematical foundations, model interpretability, formal reasoning, and AI governance. His ICLR 2026 papers include Temporal Sparse Autoencoders (oral) and GradPCA, addressing interpretable representations and reliable detection of distribution shifts. He coauthored ProofCompass (AI for MATH workshop at ICML 2025), which combines language-model guidance with specialized theorem provers, and the preprint An Instrument to Evaluate Governance Proposals: AI Policy Analysis at Scale. He is also a member of the European Commission's Scientific Panel of Independent AI Experts. This combination of technical research and policy engagement would help the organizing team connect proof and evaluation standards with institutional accountability, the interpretation of scientific claims, and the governance of increasingly capable research systems.

Role: Call for contributions and community outreach, preparing and distributing the CFP and reaching relevant research communities to attract a diverse range of contributors.

Website: Personal homepage | Google Scholar: Profile

Publications

Anja Šurina

PhD Candidate; Research Fellow

EPFL / Inherent Laboratories

Anja Šurina is a PhD candidate at EPFL, advised by Caglar Gulcehre and Emmanuel Abbe, and a Research Fellow at Inherent Laboratories. She studies post-training and test-time scaling for scientific and mathematical discovery, and previously worked on Google DeepMind's AlphaProof Nexus team. Her publications include Algorithm Discovery With LLMs: Evolutionary Search Meets Reinforcement Learning (COLM 2025), Advancing Mathematics Research with AI-Driven Formal Proof Search (Science 2026), and Training AI Scientists to Replicate Research. Her experience connects the construction of discovery systems with questions about formal verification, reproducibility, and research innovation.

Role: Contributed program chair, overseeing peer review, scientific evaluation, reviewer discussions, and contributed talk selection, with particular attention to discovery systems and reproducibility.

Website: Personal homepage | Google Scholar: Profile

EPFL profile

Siqiao Mu

PhD Candidate in Engineering Sciences and Applied Mathematics

Northwestern University

Siqiao Mu is a PhD candidate in Engineering Sciences and Applied Mathematics at Northwestern University, advised by Diego Klabjan. Her research addresses problems in nonconvex optimization, reinforcement learning, machine unlearning, and AI safety and privacy. Her work includes On the Convergence Rate of LoRA Gradient Descent (ICML 2026, oral) and Rewind-to-Delete: Certified Machine Unlearning for Nonconvex Functions (NeurIPS 2025). She served as president of Northwestern's SIAM student chapter from 2023 to 2025 and has experience organizing research meetings and reviewing for ICML and NeurIPS. She brings mathematical and machine-learning expertise to the assessment of theoretical claims and contributed work.

Role: Author communications and contributed presentation coordination, managing submission queries, decision notifications, accepted-paper requirements, and the scheduling and preparation of contributed talks.

Website: Personal homepage | Google Scholar: Profile | CV

Leo Shaposhnik

PhD Researcher in Theoretical Physics

Freie Universität Berlin

Leo Shaposhnik is a final-year PhD researcher in theoretical physics at Freie Universität Berlin, working with Alexander Jahn and Jens Eisert on quantum information, holography, and emergent spacetime. His publications include On Infinite Tensor Networks, Complementary Recovery and Type II Factors and Searching for Emergent Spacetime in Spin Glasses. He develops and uses AI agents for theoretical exploration, adversarial review, scientific computing, and formalization in Lean. His research statement examines the reliability of AI-assisted calculations and proofs, the role of human verification, and how publication incentives shape scientific work. He co-organizes a weekly seminar on quantum many-body physics and has experience leading distributed research collaborations. He brings a theoretical physicist's practical perspective on checking AI-generated arguments and connecting computational progress to scientific understanding.

Role: OpenReview and submission operations, configuring OpenReview and supporting submission workflows and access to workshop materials.

Website: Personal homepage | Publications: INSPIRE profile

What we hope to achieve

Contributions will examine claims of novelty, creativity, and discovery through arguments, analysis, and empirical evidence.

We aim to clarify where existing scientific standards remain adequate, where they may need revision, and which requirements depend on the discipline or type of contribution. For the ICLR community, this includes identifying what would make evaluations of scientific AI more informative about novelty, validity, human involvement, and usefulness to subsequent research. The intended outcome is a better-grounded agenda for research and practice, with the reasoning behind both agreement and disagreement made explicit.

SCOPE brings together communities with different methods, standards of evidence, and views about AI's role in research. Our organizing team includes faculty, industry researchers, and doctoral researchers, while the confirmed speakers contribute perspectives from mathematics, machine learning, cosmology, biomedicine, and ecology.

These decisions directly concern the ICLR community. They influence how research agents are trained and evaluated, what counts as successful scientific assistance, and which human contributions remain visible in reported results. Physical grounding and experimental validation, emphasized by Anandkumar (2026), illustrate why the generation of plausible ideas alone cannot establish scientific progress. SCOPE will examine how standards for contribution should inform model evaluation, disclosure of human involvement, and the evidence used to substantiate discovery claims.

Topics we hope to cover

We will invite contributions on what should count as scientific contribution when AI participates in research. The workshop takes mathematics as its starting point, with contributions from other scientific disciplines and machine learning providing complementary tests of shared principles. We welcome clear arguments, contrasting perspectives, and evidence that can advance discussion across these communities.

Contributions may address one or more of the following topics:

Scientific contribution, originality, and creativity: distinguishing novelty, significance, explanation, and useful synthesis.
Evidence for discovery: verification, replication, prior-art assessment, data leakage, and the disclosure of human intervention.
Human understanding and responsibility: the relationship between correctness, explanation, oversight, and accountability.
Authorship and credit: attribution to humans and AI systems, contribution statements, and recognition for validation and interpretation.
Provenance and publication: responsible release, independent review, and responsibility for verifying, explaining, and correcting AI-generated results.
Research incentives, autonomy, and access: who sets research priorities, how validation and interpretation are rewarded, and how scientific communities can participate independently of frontier AI laboratories.

Contributions may include:

Evaluations of novelty, creativity, and scientific discovery that distinguish new findings from retrieval, recombination, or benchmark optimization.
Position papers and think pieces that develop a well-supported argument about scientific contribution, originality, provenance, or standards of evidence.
Studies of self-improving agents for scientific discovery, including whether improvements transfer to new scientific problems and how human input and resource use affect the results.
Studies of tools and environments for scientific experimentation, including how they support reproducibility, independent verification, and documentation of human and AI contributions.
Comparative studies of human, AI-assisted, and autonomous research workflows, including negative results and failures that reveal limits of current evaluations.

The call for contributions will open on November 29, conditional on acceptance.

Get involved

We’re looking for reviewers and area chairs from mathematics, the sciences, and machine learning, as well as sponsors who can help make SCOPE accessible to more participants. If the workshop’s questions connect with your work, we’d love to hear from you.

If the workshop is accepted, reviewing will run from late January to mid-February 2027, with no more than five papers per reviewer.

Reviewers: We welcome researchers at every career stage, including PhD students who have published in a related area. We’ll match submissions to your expertise and availability.
Area chairs: We’re looking for faculty and senior researchers to oversee a small set of submissions in their area, read the reviews, contribute their own assessment, and recommend decisions. We especially welcome expertise in mathematics and formal verification, the experimental and observational sciences, AI evaluation and reproducibility, and the philosophy and sociology of science.
Sponsors: We’d be happy to hear from organizations interested in supporting travel and registration for students and researchers with limited institutional funding. Your support would help people take part who might otherwise be unable to attend.

Express interest

References

Advisory Group on Mathematics and Artificial Intelligence. Responsible release of AI-generated mathematics. September 29, 2026a. URL https://agmai.org/general-sep29/.
Advisory Group on Mathematics and Artificial Intelligence. On OpenAI's release of mathematical results, October 2026b. URL https://agmai.org/. October 6.
Anima Anandkumar. How do we build AI to push the frontiers of scientific discovery? Dædalus, 155(1–2):216–228, 2026. doi:10.1162/DAED.a.983. URL https://www.amacad.org/publication/daedalus/how-do-we-build-ai-push-frontiers-scientific-discovery.
Association for Human Mathematics Communications Working Group. AHM statement on OpenAI's October 6 release of mathematical documents, October 2026. URL https://terrytao.wordpress.com/2026/10/07/ahm-statement-on-openais-october-6-release-of-mathematical-documents/. October 7. Guest statement republished on Terence Tao's blog.
Sara Beery. AI & ecology: From tool to transformation. Dædalus, 155(1–2):364–367, 2026. doi:10.1162/DAED.a.996. URL https://www.amacad.org/publication/daedalus/ai-ecology-tool-to-transformation.
Juraj Gottweis, Wei-Hung Weng, Alexander Daryin, Tao Tu, Anil Palepu, Petar Sirkovic, Artiom Myaskovsky, Felix Weissenberger, Keran Rong, Ryutaro Tanno, Khaled Saab, Dan Popovici, Jacob Blum, Fan Zhang, Katherine Chou, Avinatan Hassidim, Burak Gokturk, Amin Vahdat, Pushmeet Kohli, Yossi Matias, Andrew Carroll, Kavita Kulkarni, Nenad Tomasev, Yuan Guan, Vikram Dhillon, Eeshit Dhaval Vaishnav, Byron Lee, Tiago R. D. Costa, José R. Penadés, Gary Peltz, Yunhan Xu, Annalisa Pawlosky, Alan Karthikesalingam, and Vivek Natarajan. Towards an AI co-scientist. arXiv preprint arXiv:2502.18864v1, 2025. URL https://arxiv.org/abs/2502.18864v1.
Raul Jimenez, Boris Bolliet, Francisco Villaescusa-Navarro, Rabih Zbib, Benjamin Wandelt, David N. Spergel, Thomas Meier, Jessica Montgomery, Hana Aliee, and Licia Verde. AI scientists as engines of discovery: A case for development within reformed institutions. arXiv preprint arXiv:2606.22859, 2026. URL https://arxiv.org/abs/2606.22859.
Sayash Kapoor and Arvind Narayanan. Leakage and the reproducibility crisis in machine-learning-based science. Patterns, 4(9):100804, 2023. doi:10.1016/j.patter.2023.100804.
Hiroaki Kitano. Nobel Turing Challenge: Creating the engine for scientific discovery. npj Systems Biology and Applications, 7:29, 2021. doi:10.1038/s41540-021-00189-3.
Sabina Leonelli. Philosophy of Open Science. Elements in the Philosophy of Science. Cambridge University Press, 2023. doi:10.1017/9781009416368.
Chris Lu, Cong Lu, Robert Tjarko Lange, Yutaro Yamada, Shengran Hu, Jakob Foerster, David Ha, and Jeff Clune. Towards end-to-end automation of AI research. Nature, 651(8107):914–919, 2026. doi:10.1038/s41586-026-10265-5.
Alex Meiburg. [On Comparator challenge specifications]. Oh Hai There, October 8, 2026a. URL https://blog.ohaithe.re/post/829974396980641792/id-like-to-alert-people-that-several-of-the. Untitled blog post; descriptive title supplied.
Alex Meiburg. Thoughts on AI solving “math”. Oh Hai There, October 8, 2026b. URL https://blog.ohaithe.re/post/829972087725817856/thoughts-on-ai-solving-math.
Lisa Messeri and M. J. Crockett. Artificial intelligence and illusions of understanding in scientific research. Nature, 627(8002):49–58, 2024. doi:10.1038/s41586-024-07146-0.
OpenAI. Sharing AI progress in mathematics. October 6, 2026. URL https://openai.com/index/sharing-ai-progress-in-mathematics/.
Christian Szegedy. Is mathematics over, or just graduating? Essay, October 3, 2026. URL https://docs.google.com/document/d/e/2PACX-1vTSh-pyNP3Gi99WMmsinnLmE9V5CDI0HEm6WGbIPMNt3V5SGlClHF8-BetotKHOImrvQDSXmFdiOw8D/pub.
Terence Tao. Mathematics in the age of AI. arXiv preprint arXiv:2608.16753, 2026. URL https://arxiv.org/abs/2608.16753.
Kaiyue Wen, Tengyu Ma, and Percy Liang. Priced guidance: Can language models generate future research ideas? arXiv preprint arXiv:2610.04976, 2026. URL https://arxiv.org/abs/2610.04976.

Contact

Mariya I. Vasileva · Lead organizer and designated ICLR contact