ConjectureBench: 15k open math problems for AI agents, frontier models crack decades-old conjectures
RexDouglass · x · 2026-08-01
Bespoke Labs introduces ConjectureBench, a collection of 15,000 open math problems designed for AI agents. Frontier models are already solving long-standing problems, such as disproving the Jacobian conjecture (open since 1939) with Claude Fable 5, and tackling a 100+ year old Maxwell problem.
More from Research
- GPT-4 Can Accurately Predict Social Science Experiment Outcomes, Nature Study Shows — RobbWiller · 2026-08-01
- OpenAI's GPT-Red: Self-Play Agent Boosts Prompt Injection Defense to 95.9% — alex_verem · 2026-08-01
- Researcher Explores 'Load-Bearing Tokens' Created by LLM Training Optimization — tokenbender · 2026-08-01
- LatchBio on AI in Biology: Frontier Models Untrustworthy Without Verifiable Benchmarks — AI Engineer · 2026-08-01
- LabEvolver: Robot Scientist Hits 91% Success Rate in Continuous Tasks — imjustnewatai · 2026-08-01
- AI Solves 2-Year-Old Math Conjecture in Minutes with Legible Proof — abeirami · 2026-08-01