Last Verifier Wins the Format, Verifier Set Wins the Score: Decoupling Surface Style from Capability in Sequential RLVR

Reinforcement learning with verifiable rewards (RLVR) is a central recipe for improving language-model reasoning, but composing heterogeneous verifiers remains under-specified. A natural approach is to optimize multiple verifiers jointly, yet this can couple reward signals with different token-level pressures, such as strict instruction-following constraints and final-answer correctness. We study a simpler alternative: staged verifier composition. In a controlled Llama-3.1-8B post-training study, we compare single-objective RLVR, naive joint RLVR, and two sequential pipelines over instruction-following and math verifiers. Sequential RLVR improves average accuracy across IFEval, GSM8K, and HumanEval over the joint baseline by more than three points, while producing substantially shorter GSM8K generations. Reversing the two stages yields nearly identical aggregate task accuracy, but measurably changes surface format. These results support a capability-style decoupling view: the verifier set is more predictive of benchmark capability, whereas the last verifier disproportionately shapes output format. We frame sequential verifier composition as a lightweight diagnostic approach for efficient reasoning-oriented post-training.