OpenFrontiers: A Collection of Open Research Problems and a Benchmark for Scientific Rediscovery
Summary
- Researchers introduce OpenFrontiers, a cross-disciplinary resource containing a structured corpus of 12,151 open research problems.
- The resource features a retrospective benchmark of 200 mathematical problems, including 110 resolved in 2026 and 90 open ones.
- Each benchmark problem includes a Lean 4 statement checked by automated probes and independent reviews.
- Evaluating seven LLM agents on the HQ-50 subset showed that closed-book rediscovery is rare, while live-web success largely reflects retrieving published solutions.
Why it matters
OpenFrontiers provides a standardized evaluation framework to measure research-level problem solving and scientific rediscovery in large language models.