New paths through mathematics.
FunSearch combined language models with program evaluation to discover mathematical constructions.
Original sources ↓ · Revision history ↓
Demonstrated · source published 2023-12-14
The human problem
Some combinatorial problems have vast search spaces.
The prior constraint
Researchers develop and evaluate candidate constructions.
AI’s actual role
Generated programs were tested, selected, and iteratively improved.
The documented result
The study found new constructions for the cap set problem and useful packing heuristics.
Why it may matter
Researchers gained an additional experimental tool.
Limitations
The approach depends on an evaluator; it is not a general proof engine.
Unresolved questions
Apply verifiable search to other suitable problems.
Source history & evidence assessment
- Maturity
- Demonstrated
- Claim confidence
- unassessed
- Event date
- Not recorded
- Source published
- 2023-12-14
- Captured
- Not recorded
- Last source review
- 2026-09-05
- Editorial method
- Original source check
- Place / relevance
- London, United Kingdom · institution-location
Legacy source check; no named human reviewer is recorded in this projection.
Maturity describes the tested or operational setting. Confidence describes support for the particular claim; one does not determine the other.
Original sources
Mathematical discoveries from program search with large language models ↗ · paper
Institutions: Google DeepMind
Explore the underlying question
Related developments
Editorial connections between distinct settings and results; these links do not imply replication.
Revision & correction history
No corrections recorded.