Many valuable tasks lack an automatic way to check if an answer is right. Writing a scientific abstract or grant proposal is an example. How can a language model train on these if there is no explicit verifiable reward?
Reinforced Adversarial Learning trains a judge to tell apart high quality reference examples from the model outputs. The generative model is trained to produce outputs that fool the judge. The judge confidence becomes the reward. Only example answers are needed, not reasoning traces or verifiers.
Across math, code, and chemistry, this matched training with explicit verifiers. Generating grant abstracts for the National Institutes of Health, the outputs fooled human scientists and LLM judges more often than frontier models. This allows training reasoning models wherever good answers can be collected but not easily scored.