Reinforcement‑learning (RL) post‑training markedly improves the reasoning ability of large language models, yet its dominant cost stems from autoregressive rollout generation, often exceeding 70% of the total training time. Existing speculative‑decoding approaches are ill‑suited for this scenario: model‑based draft generators consume scarce GPU memory and can become misaligned as the policy evolves, while model‑free methods fail to exploit the repetition across training epochs.\ \ We observe that recent rollouts of the same prompt provide a high‑quality, self‑updating draft source. Leveraging this insight, we introduce NgramRL, a per‑prompt n‑gram drafter built from a bounded recent rollout history. NgramRL requires no auxiliary model, drafter training, or persistent GPU‑resident drafter state, and target‑model verification preserves the rollout distribution.\ \ To hide CPU cache lookup and scheduling overhead, we further propose an adaptive fan‑out prefetcher that predicts likely verification outcomes $(j, b)$, where $j$ is the accepted draft length and $b$ is the continuation token emitted by the target model. A causal budget allocator concentrates prefetches on high‑probability outcomes, while uncovered cases fall back to sequential lookup without altering committed tokens.\ \ Across two model families and three reasoning and structured‑generation workloads, NgramRL commits 3.60–7.77 tokens per target‑model forward in the warm‑cache regime, achieving up to 1.99× the throughput of SuffixDecoding and 7.0× that of prompt‑lookup decoding. On the Spider benchmark, the prefetcher covers 96.6% of Qwen3‑8B outcomes with a candidate budget of 16, improving measured rollout throughput by 1.38× over synchronous NgramRL and approaching the 1.40× overlap ceiling.\ \ Review: NgramRL exploits n‑gram statistics from recent rollouts to enable efficient speculative decoding without extra models, and its adaptive prefetcher further reduces CPU overhead, delivering substantial speed‑ups in structured generation tasks.