Researchers at UC Berkeley and AWS AI Labs trained an AI coding agent to use a program verifier’s detailed feedback, rather than only a final pass-or-fail result. Their Oct. 13 preprint reports that more generated code passed formal verification in two benchmarks after training; the paper is under review.
Verifier diagnostics turn failure into partial credit
A program verifier checks whether code satisfies a formal specification: a statement of what a program must guarantee. Dafny, one of the systems used in the paper, lets developers write such specifications alongside code and check them with a verifier, as described in the Dafny reference manual. The researchers’ agent asks a live verifier which requirements the code has not yet shown it meets, then revises the implementation or the proof annotations that explain why it meets the specification.
Rather than giving every failed attempt the same zero score, training awards partial credit based on the weighted share of the instance’s fixed proof requirements that an attempt satisfies. An integrity audit penalizes changing the original specification or adding an unproved assumption. A separate proof-specialist model, trained beforehand to repair proof annotations and kept fixed during code-generation training, can replace the agent’s annotations while leaving its implementation unchanged.
Reported pass rates rose on code and patch tasks
For Qwen3.6-35B-A3B, researchers measured code-generation tasks held out from training. On DafnyBench, 66.1% of the starting model’s submissions were accepted by both the verifier and integrity audit; the selected verifier-grounded reinforcement-learning checkpoint reached 93.5%. On SWE-Proof, the corresponding rates were 56.3% and 100%.
In a separate held-out SWE-Proof repository-patch task at the same model scale, the setup with a proof specialist included saw the share of patches passing each project’s withheld tests rise from 59.6% for the starting model to 72.5% for the selected trained checkpoint. The authors selected each run’s checkpoint for its highest held-out verification rate, so those evaluation results also influenced which checkpoint they reported. The preprint is under review, and a verifier only checks a program against the specification it is given.
Sources
- George Ma et al., “Verifier-Grounded Agentic RL for Provably Correct Code,” arXiv:2610.19298v1 (Oct. 13, 2026; preprint under review)
- Dafny Reference Manual, program specifications and verification
Image credit: Slashme, RustCodeOnScreen.jpg, CC0 1.0. Illustrative code photo; it does not depict the system studied.
