Supporting the forming of conjectures rather than producing answers — about $1.08M to UCLA (mathematical and physical sciences code)
An award developing a human-centred, AI-assisted framework supporting the iterative process by which mathematicians form conjectures, refine definitions, explore partial arguments and validate results. Collaboration paradigms are to adapt to user expertise while preserving interpretability and mathematical intent. The budget code is 47.049, covering mathematical and physical sciences.
Grant overview (primary data)
- Award amount$1,080,001
- RecipientUniversity of California-Los Angeles (California)
- ProgramMSPA-INTERDISCIPLINARY
- Period2026-09-01 〜 2029-08-31
- FunderU.S. National Science Foundation (NSF) / NSF
Key points
- Supports the iterative process of forming conjectures, refining definitions, exploring partial arguments and validating results.
- Collaboration paradigms are to preserve interpretability and mathematical intent while adapting to user expertise.
- Methods named include curriculum learning, contrastive learning and research-aligned synthetic data generation.
- Interactive reasoning tools are built using formal theorem proving.
- The budget code 47.049 accounts for 5 of the 120 NSF awards this site holds as of 2026-09-01, a small number.
- The tool supports the process — formulating conjectures, refining definitions, exploring partial arguments — while preserving interpretability and intent.
1The subject is the process, not the answer
Talk of AI in mathematics gathers around whether hard problems can be solved. That is not what this project addresses. What the abstract names is an iterative process: formulating conjectures, refining definitions, exploring partial arguments, validating results.
Research in practice does not end when one right answer appears; it advances while what to ask is decided again. Intent-driven is the phrase for that quality. Placing the object of support on the process changes how the tool is designed.
2Interpretability and intent are written in as conditions
On collaboration paradigms the abstract names two conditions: preserving interpretability, and preserving mathematical intent. Alongside them, adapting to user expertise.
A result that is correct but cannot be followed is unusable as mathematics. Preserving intent means the tool does not take away the direction the user is heading in. Adapting to expertise allows for a beginner and a researcher needing different support.
3The budget code is mathematical and physical sciences
The budget code is 47.049, covering mathematical and physical sciences. Among the 120 NSF awards this site holds as of 2026-09-01 it accounts for 5, a small number. The largest is 47.076 for education at 26, followed by 47.084 for technology and innovation at 23 and 47.070 for computing at 21.
The methods named are curriculum learning and contrastive learning, high-quality research-aligned synthetic data generation, and interactive tools built on formal theorem proving. Capturing on the model side the diversity of proof styles and definition systems mathematicians use is part of the aim. The period runs three years, from 2026-09-01 to 2029-08-31.
4What a tool supports changes how it is built
Talk of AI in mathematics gathers around whether hard problems can be solved. That is not what this project addresses. What the abstract names is an iterative process: formulating conjectures, refining definitions, exploring partial arguments and verifying results.
The forms of collaboration the abstract names are preserving interpretability and preserving mathematical intent. The methods listed are curriculum and contrastive learning, generation of high-quality research-grade synthetic data, and interactive tools using formal theorem proving — an attempt to capture, on the model side, how varied proof styles and systems of definition are across mathematicians.
Why it matters
Whether a tool produces answers or supports the thinking changes the design requirements considerably. Preserving interpretability and the user intent is asked of tools used by experts generally. So is varying the support with the expertise of the user, and neither is peculiar to mathematics.
FAQ
Is this an AI that solves problems?
Why is interpretability a condition?
What is formal theorem proving?
Sources (primary)
Source: NSF Award Search (U.S. National Science Foundation, public domain). Amounts are the obligated amount. For privacy, we do not handle principal investigator names.
- NSF Award (original, official)
- NSF Award ID: 2617441