New hybrid approach combines LLMs with proof sketches to boost automated theorem proving accuracy

Automating the creation of formal mathematical proofs remains a significant challenge for AI systems, held back by the sheer effort required to construct rigorous, correct arguments manually. Researchers have now developed a new strategy that bridges two previously separate approaches to machine-assisted theorem proving, demonstrating that smaller language models can achieve stronger results when given the right architectural support.

How the hybrid approach works#

HybridProver represents a shift in how LLMs tackle formal verification problems. Rather than committing entirely to either generating complete proofs in one shot or building them step-by-step through tactical reasoning, the framework uses proof sketches as a middle ground. This intermediate representation allows the system to preserve the useful structure of partially correct proofs while combining the benefits of both high-level planning and detailed logical steps.

The researchers implemented their method in Isabelle/HOL, a widely-used formal verification language, and trained two 7-billion-parameter language models on curated datasets optimized for this task. The key insight is that smaller, efficiently-trained models can outperform larger systems when equipped with a better framework for organizing reasoning.

Impressive benchmark results#

On the miniF2F Isabelle benchmark—a standard test suite for theorem-proving systems—HybridProver achieved a 73.8% success rate, surpassing the previous state-of-the-art at 61.9%. This improvement is particularly notable because it was obtained without scaling up to very large LLMs; instead, the architectural innovation did the heavy lifting.

The work included systematic ablation studies examining how dataset quality, training setup, and sampling methods impact performance, providing insights into which design choices matter most for proof generation systems.

Why this matters#

This research suggests that automated formal verification—a critical tool for ensuring reliability in safety-critical systems—may become more accessible and efficient. By showing that compact models can handle complex mathematical reasoning through smarter integration of proof synthesis strategies, the work opens pathways for deploying theorem-proving AI in resource-constrained settings.