Reinforcement learning for code typically rewards models for passing tests. The problem, according to a new arXiv paper, is that a model can pass those tests while still being wrong on inputs the tests do not cover. Test-based rewards therefore do not guarantee that a program is actually correct.

The paper proposes grounding agentic reinforcement learning in a program verifier instead. A program verifier decides whether a program satisfies a formal specification, offering a stronger signal than test outcomes. By using verifier feedback to guide the agent's learning, the approach aims to produce code with provable correctness rather than mere test-passing behavior.

The source is a single arXiv abstract, so the article reflects only that summary. It does not include experimental results or comparisons to other methods, and no independent evaluation is reported.