A Relative-Budget Theory for Reinforcement Learning with Verifiable Rewards in Large Language Model Reasoning
This work addresses the challenge of optimizing RL for large language model reasoning, which is incremental as it builds on existing RL paradigms to explain and improve efficiency.
The paper tackles the problem of varying effectiveness of reinforcement learning (RL) in improving large language model reasoning across different tasks and compute budgets by proposing a relative-budget theory, showing that a relative budget parameter ξ determines sample efficiency and identifying an optimal range ξ ∈ [1.5, 2.0] for maximizing learning efficiency and reasoning performance.
Reinforcement learning (RL) is a dominant paradigm for improving the reasoning abilities of large language models, yet its effectiveness varies across tasks and compute budgets. We propose a \emph{relative-budget} theory explaining this variation through a single quantity called relative budget $ξ:= H/\mathbb{E}[T]$, where $H$ is the generation horizon (token budget) and $T$ denotes the number of tokens until the first correct solution under a base policy. We show that $ξ$ determines sample efficiency by controlling reward variance and the likelihood of informative trajectories. Our analysis reveals three regimes: in the \emph{deficient} regime ($ξ\to 0$), informative trajectories are rare and the sample complexity explodes; in the \emph{balanced} regime ($ξ=Θ(1)$), informative trajectories occur with non-negligible probability and RL is maximally sample-efficient; and in the \emph{ample} regime ($ξ\to \infty$), learning remains stable but marginal gains per iteration diminish. We further provide finite-sample guarantees for online RL that characterize learning progress across these regimes. Specifically, in a case study under idealized distributional assumptions, we show that the relative budget grows linearly over iterations. Our empirical results confirm these predictions in realistic settings, identifying a budget $ξ\in [1.5, 2.0]$ that maximizes learning efficiency and coincides with peak reasoning performance.