Compile to Compress: Boosting Formal Theorem Provers by Compiler Outputs

A team of researchers has published a paper titled 'Compile to Compress: Boosting Formal Theorem Provers by Compiler Outputs,' introducing a novel method to make AI-driven formal theorem proving significantly more efficient. The core innovation exploits a key observation in formal verification: compilers act as a powerful compression mechanism, mapping a vast, diverse space of failed proof attempts down to a much smaller, structured set of error messages or failure modes. By leveraging this inherent structure, the researchers' 'learning-to-refine' framework allows AI models to perform targeted, local corrections based on explicit feedback from a verifier, rather than brute-forcing solutions through massive roll-outs or extremely long context windows.

This approach directly tackles the major scalability bottleneck in current Large Language Model (LLM) applications for theorem proving, which is prohibitive test-time compute. Instead of accumulating a lengthy and costly history of proof attempts, the system performs a more efficient tree search that conditions each step on the verifier's output. Extensive evaluations demonstrate that the method consistently amplifies the reasoning capabilities of base provers across different model scales.

Notably, the framework achieved state-of-the-art performance on the challenging PutnamBench benchmark among publicly reported models with approximately 8 billion and 32 billion parameters, all while operating under comparable test-time compute budgets. This represents a scalable new paradigm for 'verifier-guided reasoning,' where the AI's exploration is tightly coupled with and constrained by the formal feedback from a compiler or proof assistant, leading to more reliable and efficient automated reasoning systems.