ResearchResearch paperReinforcement Learning · Efficiency & Inference · Large Language Models1 source · Oct 13, 2026

Verifier-Grounded Agentic RL for Provably Correct Code

Reinforcement learning for code usually rewards passing tests, which a model can do while remaining wrong on the inputs the tests do not cover.

A program verifier instead decides whether a program satisfies a formal specification for all inputs.
Language models are almost always paired with verifiers at inference time only, and turning a verifier into a reward raises two difficulties.
We build an agentic environment and reward that address both.
Across Qwen3 models from 1.7B to 14B and a 35B Qwen3.6 mixture-of-experts model, held-out verification on code generation rises from 66.1% to 93.5% on DafnyBench and from 56.3% to 100% on SWE-Proof.

[1]Verifier-Grounded Agentic RL for Provably Correct Code
arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 11:17 PM
Reinforcement learning for code usually rewards passing tests, which a model can do while remaining wrong on the inputs the tests do not cover.
A program verifier instead decides whether a program satisfies a formal specification for all inputs.

Extractive summary: sentences quoted from the sources.

Before this