# Train at Moving Edge: Online-Verified Prompt Selection for Efficient RL Training of Large Reasoning Model

Source: [Train at Moving Edge: Online-Verified Prompt Selection for Efficient RL Training of Large Reasoning Model](<https://www.sota2.com/research/paper/69c60eff0636e5bf4afc84e3>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-06-10T05:00:46.690Z

Authors: Jiahao Wu, Ning Lu, Shengcai Liu, Kun Wang, Yanting Yang, Bailong Lin, Chen Jason Zhang, Li Qing, Ke Tang

Published: 2026-03-26T08:52:35.000Z

## Abstract

Reinforcement learning (RL) has become essential for post-training large language models (LLMs) in reasoning tasks. While scaling rollouts can stabilize training and enhance performance, the computational overhead is a critical issue. In algorithms like GRPO, multiple rollouts per prompt incur prohibitive costs, as a large portion of prompts provide negligible gradients and are thus of low utility. To address this problem, we investigate how to select high-utility prompts before the rollout phase. Our experimental analysis reveals that sample utility is non-uniform and evolving: the strongest learning signals concentrate at the \`\`learning edge", the intersection of intermediate difficulty and high uncertainty, which shifts as training proceeds. Motivated by this, we propose HIVE (History-Informed and online-VErified prompt selection), a dual-stage framework for data-efficient RL. HIVE utilizes historical reward trajectories for coarse selection and employs prompt entropy as a real-time proxy to prune instances with stale utility. By evaluating HIVE across multiple math reasoning benchmarks and models, we show that HIVE yields significant rollout efficiency without compromising performance.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Mathematical Reasoning on Math500, AIME24, AMC, Gaokao, Minerva Math, and Olympiad Bench](<https://www.sota2.com/research/sota/mathematical-reasoning-on-math500-aime24-amc-gaokao-minerva-math-and-olympiad-bench.md>): Mathematical Reasoning; Math Reasoning Suite (Math500, AIME24, AMC, Gaokao, Minerva Math, Olympiad Bench); Math500 Score: 93.

- [Mathematical Reasoning on DAPO + MATH](<https://www.sota2.com/research/sota/mathematical-reasoning-on-dapo-plus-math.md>): Mathematical Reasoning; DAPO + MATH; Training Time (hours): 10.1.

- [Mathematical Reasoning on OpenR1 Math Evaluation Suite (Math500, AIME24, AMC, Gaokao, Miner., Olymp.) (test)](<https://www.sota2.com/research/sota/mathematical-reasoning-on-openr1-math-evaluation-suite-math500-aime24-amc-gaokao-miner-olymp-test.md>): Mathematical Reasoning; OpenR1 Math Evaluation Suite (Math500, AIME24, AMC, Gaokao, Miner., Olymp.) (test); Math500 Accuracy: 82.6.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2603.25184>)

[PDF](<https://arxiv.org/pdf/2603.25184.pdf>)
