On November 12, researchers from GAIR and partner Chinese institutions published daVinci-Prover, an approach to training and operating a mathematical AI agent. Kaldata covered the work on November 14. The noteworthy combination is proof generation with an external verifier. The authors report 97.1% success on MiniF2F with memory compaction. That is a result from a particular experiment, rather than a score for mathematical problems in general.

What the verifier actually checks

The agent proposes steps, while Lean checks whether the formal result follows from the specified statement and assumptions. Compaction shortens older history while retaining the current theorem and proof. Our review did not establish independent reproduction of the trained models.

For readers, a proposed solution and an accepted proof are materially different. A language system can produce fluent reasoning that appears complete. Machine checking supplies a separate acceptance criterion. However, it checks the formal statement it receives: successfully proving an incorrectly translated statement does not fix the mismatch with what a person intended to ask.

MiniF2F itself contains formalized exercises drawn from competitions and classroom mathematics. Its maintainers ask researchers to identify the version or commit used because statements receive corrections. Comparing percentages consequently requires comparable tasks and settings. Reading a natural-language problem, translating it into a formal statement and proving that statement are separate parts of the overall workflow.

Code and trained models are different artifacts

The public repository supplies a Lean environment, local premise retrieval and final acceptance through Comparator. Its current example uses DeepSeek with its own interaction budget. Those tables should not be merged with the Qwen training results in the paper. Available code makes the workflow inspectable; it does not establish that every announced model weight and training dataset has already been released. Code availability needs to be assessed separately from trained-model availability.

Lean's documentation provides another reason to inspect assumptions: the system supports axioms, and the chosen foundations affect what a proof means. A useful review therefore looks beyond a successful verification message to the precise theorem, dependencies and environment used to produce it. This is especially important when results move between experimental harnesses or library versions.

Our editorial interpretation is practical. daVinci-Prover illustrates how an agent can receive checkable feedback from a specialized tool. Applying that approach beyond the published tests calls for repeatable runs with fixed versions, records of unsuccessful attempts and a separate review of the formalization. These requirements define what the result supports; a high benchmark percentage alone cannot establish reliable performance on an unseen mathematical workflow.

An organization considering a trial should first decide which part it needs help with: proposing a statement, searching existing mathematics or checking an already formalized proof. Those tasks have different failure modes. Keeping them distinct makes it possible to assess whether the agent actually reduces work without quietly shifting errors into a stage the benchmark did not measure.

The cover is an AI-generated editorial illustration, not a documentary photograph.