Beyond the answer
Can AI Propose Interesting Research Questions?
In Douglas Adams’s The Hitchhiker’s Guide to the Galaxy, a supercomputer finds the ultimate answer: 42. The question itself remains unknown. After42 moves from answering questions to finding questions worth asking.
Frontier AI systems have made striking progress on solving research-level problems in mathematics, statistics, theoretical computer science, and physics. Research is broader than solving well-defined problems, and posing interesting questions is one important part of it. We envision AI and scientists shaping research directions together. After42 studies whether AI can propose research questions that domain experts find genuinely interesting, by developing each question through attempts to solve it. Our long-term goal is to help AI and scientists interact more effectively by tracking and assessing AI capabilities in research, and to build a more comprehensive understanding of mathematics and other scientific fields. We are extending After42 to more fields and invite mathematicians and other scientists to reach out and contribute their expertise, by setting goals in their subfields and assessing the questions it proposes.
News
[Oct. 22] We release the first batch of expert-verified research questions from After42, browsable by field. Each question is published together with the expert's assessment.
Beyond solving
Research is broader than solving given problems. It develops concepts and theories, creates methods and tools, and explores new directions. Posing research questions is one important part of it, because questions shape the direction of research and where researchers invest their effort. Novel and difficult questions are relatively easy to generate, but genuinely interesting ones are rare and hard to judge. We study whether AI can propose questions that domain experts find genuinely interesting and whether AI can judge which questions are interesting.
Developing questions by solving them
The insight needed to formulate a strong question often emerges only through investigation. In mathematics, conjectures and concepts are repeatedly revised in the course of attempted proofs. Our multi-agent workflow, Agent42, builds on this idea. Starting from a researcher-written subfield goal, it generates a broad pool of literature-grounded candidate questions and filters out weaker ones. It then develops the surviving questions through rounds of exploratory solving and revision, with model judges evaluating the proposed changes. Domain experts assess the final questions in subfields of mathematics, statistics, theoretical computer science, and physics.
What we find
Experts judge some of the final questions genuinely interesting, and development trajectories show how exploratory solving sharpens questions or redirects them toward a different research target. Model judges cannot yet fully replace expert research taste. Their preferences are reproducible and steerable, but their rankings do not consistently prioritize the questions experts find genuinely interesting. See our blog post for more findings.
Building understanding together
Solving problems and posing questions are entangled in research. Our long-term goal is to help AI and scientists interact more effectively by tracking and assessing AI capabilities in research, and to build a more comprehensive understanding of mathematics and other scientific fields. Posing good questions requires this understanding, and solving open questions alone cannot build it. We release the expert-verified questions on this site as invitations for researchers to pursue, revise, or challenge.
FAQ
Who wrote these questions?
Agent42 proposed and developed each question toward a goal written by researchers in the subfield. Every published question has been verified by an expert, and we publish each question with the expert's assessment.
Can I just take one and work on it?
Yes, that is the point of publishing them. No permission needed. If you make progress on a question or solve it, tell us and we will link your work from the question card and update its status. If your work builds on a After42 question, please cite it (see below).
What do the status labels mean?
Following the convention of the Erdős problems collection, every question carries one of three statuses, open (no progress known to us), partial (meaningful partial results exist, linked from the question card), and solved (resolved, with a link to the paper or solution).
How do I cite a question?
Each question card carries a permanent link (the § next to its title) and a ready-made BibTeX entry under “cite this question”. If your work uses several questions or the collection as a whole, please also cite the project; the project citation is on the contribute page.
How can I contribute?
We would love to hear from mathematicians and scientists who want to contribute their expertise. You can set a goal in your subfield, assess questions and share your taste, pursue a published question, or help build the system. See the contribute page for details.
I found prior art, or an error in a published question.
Please tell us via the contribute page. That feedback guides how Agent42 proposes and develops questions, and we publish corrections together with what changed.
How do I know a question is actually open?
Each question was checked against the literature, tested by repeated solution attempts, and audited for validity before experts assessed it. An open status reflects our current knowledge of the literature, so do your own search before investing serious effort, and if you find prior art, tell us.
Why are there so few questions?
Interesting questions are rare. Novel and difficult questions are easy to generate in volume, so the measure we report is expert interest. Most candidates do not survive development, and we publish only questions that experts have assessed. The collection grows as experts assess new ones.