Kimina-Prover: Scaling Formal Reasoning Through Reinforcement Learning Search

Researchers have introduced Kimina-Prover, a novel approach that applies reinforcement learning (RL) during test time to enhance formal reasoning in large language models. The method focuses on improving proof generation in formal mathematics by allowing the model to explore and refine its reasoning steps through search at inference time, rather than relying solely on static training. This technique demonstrates significant gains in accuracy on challenging formal verification benchmarks, marking a step forward in combining RL search with large-scale formal reasoning systems.