Test-time scaling can improve LLM reasoning, but existing approaches typically adapt generation or apply answer selection after a predetermined generation schedule. The BOSS method instead studies within-instance allocation: how to distribute a hard per-instance token budget across sampling and selection in real time.
By making allocation decisions online, BOSS avoids the rigidity of fixed schedules and may use compute more efficiently. The approach is designed for settings where each instance has a strict token limit, making it relevant for practical deployments of large language models.
The paper is available on arXiv and introduces a new perspective on test-time scaling, focusing on budget-aware reasoning rather than simply increasing generation length.