Machine-checked proof: Lean, AlphaProof and open provers

Machine-checked proofs ensure correctness by verifying logical derivations through computational kernels. However, traditional proof assistants require human-generated proofs, limiting scalability. Neural theorem proving introduces automation, leveraging large language models to generate proofs that are then validated by a proof assistant’s kernel. This shift alters the economics of formal verification by reducing reliance on scarce human expertise.

Neural Theorem Proving and Lean 4

A proof assistant like Lean 4 operates under a strict trust model: only derivations its kernel can verify are accepted. This means that the validity of a proof depends entirely on the kernel’s logic, not on the method used to generate the proof. Whether a proof is derived through human ingenuity, brute-force search, or a language model, it must meet the kernel’s standards to be accepted. This design ensures that the correctness of the final proof is guaranteed by the kernel alone, not by the process that produced it.

DeepMind’s AlphaProof demonstrates how this trust model can be leveraged with neural methods. It reached a silver-medal level on International Mathematical Olympiad problems by combining a language model with Lean 4 verification and AlphaZero-style reinforcement learning. During inference, it used test-time reinforcement learning to train on self-generated problem variants. Crucially, hallucination is structurally impossible in its outputs because Lean 4 only accepts proofs that it can independently verify.

Open-Source Provers and Scalability

The open-weights line in neural theorem proving has made these systems more reproducible and accessible. DeepSeek-Prover V1.5 (arXiv:2408.08152) achieved significant progress in Lean 4 proving with 7 billion parameters. Its successor, V2 (arXiv:2504.21801), introduced recursive subgoal decomposition and scaled to 671 billion parameters. Both versions were evaluated on the ProverBench dataset, which is hosted on Hugging Face. These models showcase how increasing model size and architectural improvements can improve proof generation.

Formal Methods in Agent Engineering

Formal methods have extended beyond traditional software verification into agent engineering. Lean4Agent (arXiv:2606.06523) applies formal models and verification techniques to agent workflows and trajectories. Similarly, Event-B agents (arXiv:2605.17475) use LLM assistance to synthesise and repair formal models. These tools demonstrate how formal verification can be integrated into the design and operation of complex systems, especially in domains involving autonomous agents.

Economic Impact of Neural Provers

The economics of formal verification have shifted due to neural provers. Previously, formal verification was constrained by the limited number of human proof-writers. Neural provers change this by turning proof search into a computational problem. As a result, systems that were not economically viable for formal verification, such as protocols, controllers, and agent runtimes, can now be considered. This expansion in scope occurs because proof search is no longer limited by human capacity but by compute and model quality.

Current Limitations

Despite advancements, current neural provers are best suited for competition-style, well-posed problems. The real-world challenge lies in specifying systems correctly, a task still requiring human involvement. A machine-checked proof of an incorrect specification is a failure mode that highlights the importance of accurate problem formulation. As lesson 14 notes, this remains the field’s most elegant failure mode.

What to take away

Neural theorem proving using Lean 4 and related systems shifts the economics of formal verification by making proof search computationally scalable. While current systems excel in specific domains, the core bottleneck remains in specification rather than proof generation. This distinction is key to understanding both the potential and the limitations of these tools.

Reference

Sources and further reading

• DeepSeek-Prover-V2 (arXiv:2504.21801)
• DeepSeek-Prover-V1.5 (arXiv:2408.08152)
• Lean4Agent (arXiv:2606.06523)

Apply it to your work