Brief

Researchers have developed a new neuro-symbolic framework called Enumerate-Conjecture-Prove (ECP) designed to tackle answer-construction problems in formal mathematics. This framework combines general large language models for proposing candidate answers with specialized prover LLMs for generating machine-checked proofs. ECP demonstrated success on benchmark datasets, formally solving a portion of answer-construction problems with admissible answers and proofs, outperforming existing LLM baselines. AI

Researchers have developed a new framework to improve the efficiency of formal theorem provers by leveraging compiler outputs. This method uses a learning-to-refine approach that exploits the compression of diverse proof attempts into structured failure modes by compilers. The system performs tree search, correcting errors based on verifier feedback to avoid the high computational costs of extensive proof exploration. Evaluations show this approach enhances prover capabilities and achieves state-of-the-art performance on PutnamBench for models around 8B and 32B parameters within comparable time budgets. AI