ExecuCritic only lets the critic's feedback reach the coder when the critic and the sandbox executor agree on the same rollout group

arXiv 2609.16604·medium signalXBlueskyLinkedInCopy link

RLVR on code collapses a whole program into one pass/fail bit, which makes credit assignment hard, and reviewer or tester roles in coding systems are usually prompted rather than trained against execution. ExecuCritic trains coder and critic jointly on the same execution rollouts: the critic predicts pass/fail plus short diagnostic feedback, and the coder consumes that signal only on rollout groups where the critic agrees with the executor. Across eight code benchmarks and two recent open backbones it beats GRPO without a critic, prompted reviewer systems and scalar reward models using fewer policy-gradient steps and fewer sandbox executions, with ablations attributing the gain to credit assignment rather than sampling budget.