Planned Test-Time Scaling with Coordinated Reasoning Paths

Test-time scaling has become a critical mechanism for improving large language model performance on challenging reasoning tasks (Wu et al., 2024; Snell et al., 2024; Brown et al., 2024), yet its effectiveness depends on whether additional samples explore complementary solution paths. Existing approaches typically rely on repeated sampling from a single policy, which can produce redundant attempts and limit gains despite large inference budgets. To address this limitation, we propose Planned Test-Time Scaling (PTTS), a generalized framework that treats test-time compute as a planned resource and jointly models the multi-branch sampling process. PTTS decomposes inference into a planner, which allocates the budget across diverse solution outlines, and an executor, which produces a full solution conditioned on each outline. This formulation enables both zero-shot and trainable instantiations. In PTTS-ZS, the planner generates diverse outlines in a single autoregressive pass; in PTTS-RL, we freeze the executor and train the planner directly against the pass@k reward using truncated execution rollouts. Across five mathematical reasoning benchmarks using the Qwen3-1.7B and Qwen3-4B model families, PTTS significantly outperforms standard repeated sampling: PTTS-ZS improves pass@64 by up to 6.7 points, while PTTS-RL further increases the total gain to as much as 13.4 points. Further analysis attributes these gains to broader coverage and greater diversity among reasoning branches, resulting in better budget efficiency on easier benchmarks and higher performance ceilings on harder ones.