Saved in:
Bibliographic Details
Main Authors: Wu, Jiahao, Lu, Ning, Liu, Shengcai, Wang, Kun, Yang, Yanting, Qing, Li, Tang, Ke
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2603.25184
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914424390942720
author Wu, Jiahao
Lu, Ning
Liu, Shengcai
Wang, Kun
Yang, Yanting
Qing, Li
Tang, Ke
author_facet Wu, Jiahao
Lu, Ning
Liu, Shengcai
Wang, Kun
Yang, Yanting
Qing, Li
Tang, Ke
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2603_25184
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Train at Moving Edge: Online-Verified Prompt Selection for Efficient RL Training of Large Reasoning Model
Wu, Jiahao
Lu, Ning
Liu, Shengcai
Wang, Kun
Yang, Yanting
Qing, Li
Tang, Ke
Machine Learning
Artificial Intelligence
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.
title Train at Moving Edge: Online-Verified Prompt Selection for Efficient RL Training of Large Reasoning Model
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2603.25184