A new testbed shows exactly which parts of training an AI code verifier are worth their cost.
Researchers built Aletheia, a controlled setup for training code verifiers, the models that judge whether AI-generated code actually works, using reinforcement learning with verifiable rewards (RLVR). They isolated three expensive ingredients: giving the verifier a thinking budget before it scores code, training it on wrong answers as well as right ones, and updating it on its own live outputs rather than static data. The best recipe changed with scale: on-policy training was the biggest lever for small verifiers, while thinking budget mattered more for large ones. Negative examples, code shown to be wrong, had a constant effect on picking the single best answer, but grew more useful for ranking multiple candidates as the verifier got bigger.
That matters because code verifiers are the quality control behind AI coding tools, and training them well has mostly been trial and error borrowed from other domains. The paper argues that on-policy training, the priciest ingredient, can be dropped entirely at larger scales without hurting accuracy, and skipping the thinking step saves money at smaller ones. The findings held up in a Best-of-N deployment test and two outside reward-model benchmarks, not just the authors' own setup.
Still, this is a benchmarking paper, not a shipped verifier. The real test is whether labs training these systems actually adopt the shortcuts, or just cite them in the next paper's related work section.