Kimina-Prover applies test-time RL to formal reasoning

Researchers have introduced Kimina-Prover, a large formal reasoning model that leverages test-time reinforcement learning search to enhance proof generation capabilities. The method applies search algorithms during inference to explore multiple reasoning paths, effectively trading compute time for higher accuracy on complex mathematical proofs. This approach moves beyond standard supervised fine-tuning by optimizing the model's ability to verify logical steps in real-time. The results suggest that scaling inference-time computation is a viable path to improving reliability in formal verification tasks.