Overview
The Numina fellowship is designed for research groups worldwide interested in accelerating their programs with AI tools specialized in mathematical reasoning. This fellowship aims at co-developing solutions to deep and ambitious research problems at the forefront of science and mathematics.
As part of the fellowship, Numina will support research teams by providing access to:
- General Numina AI tools and models for:
- Theorem proving with formal verification
- Autoformalization of advanced mathematical statements and proofs
- Algorithmic exploration of mathematical search spaces
- Numina AI research engineers and scientists to support the co-development of datasets and tailored AI tools / models.
- Computing resources to support the project objectives
The program will start with an initial 12-18 months collaboration during which Numina will commit resources equivalent to $50k-$200k, depending on the project needs. Programs will then be accelerated on a case-by-case basis depending on the potential outcomes.
Eligibility
- Open to public and private research organizations.
- Open to both projects in pure mathematics and scientific research in general.
- The products of research should be made publicly available (through scientific publications and open-source tools).
Examples of problems suitable for the fellowship
1. Pure mathematical research
The following types of problems are the most likely to benefit from the latest AI developments:
- Formal verification of conjecture proofs — Numina released the Numina-Lean-Agent, an Agentic Reasoning System for Formal Mathematics which solved all problems in Putnam 2025 and autoformalized 90% of this research article in Nov 2025.
- Recognizing latent mathematical structure (e.g. invariants, patterns, or heuristics) from data
- Open-ended conjectures and problem networks such as Erdős problems
2. General scientific research
The following class of problems "hard to search / easy to validate and scale" are good targets for SOTA AI systems:
- NP search space / complexity
- Scalable verification (low complexity, possibility to generate synthetic data)
Important information for applicants
To apply to the fellowship program, please complete this form before March 31, 2026 (midnight CET).
Participants wishing to sign a mutual confidentiality agreement before sharing information about their research should send us an email at [email protected] prior to completing the form.
Participants are strongly encouraged to attach a 1 to 2 pages document to their application describing the following:
- An overview of your research and the specific objectives you aim to achieve with Numina in a 12-18-month period.
- A description of how Numina can help accelerate your research, including the tools you are interested in using.
- Your academic resume and a presentation of the team working on the project.
Applications will be reviewed by the Numina team and scientific advisors, shortlisted teams will be invited to present their project in further detail via video conference in May.
Selected projects will be announced in July 2026 and fellowships beginning in the summer 2026.
General comments
- Researchers are invited to apply to the fellowship using their current research projects and to consider how AI might enhance their work. Projects do not need to be explicitly designed around AI.
- Please note that this is a fellowship program, not a grant. Numina provides access to resources such as staff support, compute, and tools — not direct funding.
About Numina
Numina is a global non-profit organization, headquartered in France, and an open-scientific collaboration whose mission is to foster the development of AI and human intelligence in the field of mathematics.
Numina won the AIMO 1st Progress Prize in July 2024 with its Numina Math 7B model. After this initial milestone, the team worked on several projects that resulted in the release of theorem-proving and auto-formalization models and agents (Kimina-Prover and Numina-Lean-Agent) as well as datasets of formalized problems and solutions (NuminaMath Lean 100k, CombiBench).
Numina-Lean-Agent solved all problems from the Putnam 2025 competition, matching the performance of leading closed-source systems. Beyond benchmark evaluations, it has been used in collaboration with mathematicians to formalize the Brascamp–Lieb theorem.
Learn more about our work at projectnumina.ai