What they found
The authors argue that good SI theorem provers need two things: strong math skills learned beforehand, and examples of long, step-by-step work with a proof checker. They built a 33.7B-token mid-training corpus (training done before fine-tuning). It covers breaking problems down, searching, checking one's own work, and revising. They also built the daVinci-Prover harness. It lets a model write proofs and talk to a Lean verifier (a tool that checks proofs) in an isolated workspace. With it they collected about 36K agent trajectories for supervised fine-tuning. On Qwen3-8B, the mid-training recipe gave an average gain of 13.7 points over a matched SFT-only baseline across MiniF2F, FATE-M, and ProofNet. On the Qwen3.5 series, they report strong results and near-saturating MiniF2F performance with memory compaction.
Why it matters
Formal proofs can be checked by machine, so they give a clear signal of whether SI reasoning is right. The authors say progress has been held back by the lack of a transparent data recipe. This paper offers one. The abstract does not say whether the data or code will be released, so that is unclear.
How it changes the race
If the result holds, a well-built mid-training corpus could matter as much as model size for SI provers. Smaller models like an 8B one could get much better at formal math. The abstract gives no cost figures, so the effect on cost is unknown.
Safety and control
Every proof step is checked by a Lean verifier, which gives a reliable way to confirm that the SI system's output is correct. The abstract does not discuss safety directly.
The abstract, by the authors
Advances in large language models (LLMs) are rapidly transforming formal theorem proving. Yet further progress remains constrained by the lack of a transparent and effective data recipe for systematically improving theorem-proving models. In this work, we argue that effective agentic provers require strong mathematical skill priors to ground theorem proving and agentic trajectories to enable long-horizon, interactive proof construction. To this end, we synthesize an agent-native 33.7B-token mid-training corpus covering decomposition, search, self-verification, and revision, together with verifier-grounded agentic trajectories. We further introduce the daVinci-Prover harness, which enables models to construct proofs and interact with a Lean verifier in an isolated workspace, and use it to collect approximately 36K high-quality agent trajectories for supervised fine-tuning. Our mid-training recipe yields an average improvement of 13.7 points over the matched Qwen3-8B SFT-only baseline across MiniF2F, FATE-M, and ProofNet. When applied to the Qwen3.5 series, our approach also yields strong performance across theorem-proving benchmarks and achieves near-saturating MiniF2F performance with memory compaction. The results highlight agent-native mid-training as an effective foundation for developing stronger interactive theorem-proving agents.
Quoted from arXiv:2611.16433. The research and its words belong to its authors.
