Kimina-Prover releases Kimina-Prover-72B with Test-time RL Search
Action Required
Organizations can leverage Kimina-Prover-72B to automate complex mathematical proof construction, potentially reducing the time and effort required for formal verification and research.
AI Impact Summary
Kimina-Prover is introducing a new theorem proving model, Kimina-Prover-72B, leveraging Test-Time Reinforcement Learning Search to solve complex mathematical problems in Lean 4. This model incorporates a novel lemma-enabled pattern and an error-fixing capability, achieving state-of-the-art performance (92.2% pass rate) on the miniF2F benchmark. The key innovation is the TTRL search framework, which allows the model to recursively discover, combine, and apply multiple lemmas autonomously, significantly improving proof construction efficiency and robustness.
Affected Systems
- Date
- Date not specified
- Change type
- capability
- Severity
- high