Reinforcement learning with verifiable rewards (RLVR) typically improves pass@1 but can fall behind its base model at large sampling budgets $k$, commonly interpreted as sharpening existing solutions. Because adjacent pass@$k$ estimates reuse generations, pointwise error bars cannot establish a crossover. We formulate paired inference on the RL-minus-base difference curve: a prompt-level multiplier bootstrap yields a one-sided dominance test and simultaneous confidence bands requiring both an early gain and a later loss. Among nine public RLVR checkpoints at 3k-16k token limits, only one narrowly meets this criterion before multiplicity adjustment; asymmetric truncation can create, mask, or reverse crossovers. On 1,060 deduplicated competition problems at 32k tokens, two 1.5B models exhibit certified crossovers around $k\approx10$-$60$, while two 7B models show no supported loss. A paired variance decomposition shows why adding prompts improves power more than deeper sampling. Prompts with identical base success probability diverge after RLVR, with dispersion replicating across independent generation halves and violating Lipschitz slope bounds for reasonable deterministic maps. Modeling RLVR as a Markov kernel, Jensen's inequality shows that conditional dispersion depresses pass@$k$ for $k\ge2$. Once dispersion is modeled, the conditional-mean log-odds slope is below one in every pair: RLVR improves hard prompts on average, while late-$k$ loss comes from a 2%-3% minority of base-solved prompts collapsing onto a repeated modal wrong answer and resisting prefix rescue up to 4,096 tokens. In controlled runs from one base model on 3,000 new problems, all four GRPO arms fall behind near $k\approx200$ and remain behind through $k=4096$, while SFT stays significantly ahead throughout.
AI 速览：将多模态智能体的开放式推理改为从预设自然语言候选集中选择，以提升推理效率并保持任务表现。
Multimodal agents commonly generate free-form reasoning before each action. For small models, limited model capacity can result in lengthy reasoning that provides little useful guidance for action generation while incurring substantial inference cost. To address this challenge, we introduce Selection-based Structured Reasoning (SSR), a framework that reformulates reasoning as selection instead of open-ended generation. SSR represents recurring high-level reasoning as pre-specified, reusable natural-language candidates. At each turn, the model selects from these reasoning candidates based on their likelihoods given the current context, without requiring an auxiliary task head. Using pre-specified reasoning traces enables parallel scoring, where teacher-forced prefilling computes token likelihoods concurrently within and across candidates using a shared context KV cache. We evaluate SSR on seven multimodal search benchmarks using 2B and 4B models. Across multiple reinforcement learning objectives and supervised fine-tuning, SSR delivers significant efficiency gains without sacrificing task performance. SSR achieves an average success rate competitive with leading search agents of the same scale, while reducing per-turn reasoning latency by over 90% and total per-question model inference latency by 28-54%. Project page: https://zfy0314.github.io/ssr-webpage/.