Test Case Scaling Trains a Test Generator Adversarially Against the Solver's Current Failure Modes

arXiv 2609.03955·medium signalXBlueskyLinkedInCopy link

RL for code generation depends on test cases that are both sound and discriminative, and those are scarce. TCS reframes test generation as an adversarial RL problem where the generator must produce counterexamples targeting the solver's present failure modes, training from a rolling policy-aligned buffer in two stages: stage 1 generates tests consistent with the reference solution, stage 2 restricts the buffer to current failure modes and learns counterexample tests. Across TACO and LiveCodeBench it improves both pass@1 and inference-time answer selection, and the learned generator also works for selecting among outputs from other LLMs.