Reinforce-Ada: An Adaptive Sampling Framework under Non-linear RL Objectives

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Xiong, Wei, Ye, Chenlu, Liao, Baohao, Dong, Hanze, Xu, Xinxing, Monz, Christof, Bian, Jiang, Jiang, Nan, Zhang, Tong
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866911303254147072
author Xiong, Wei
Ye, Chenlu
Liao, Baohao
Dong, Hanze
Xu, Xinxing
Monz, Christof
Bian, Jiang
Jiang, Nan
Zhang, Tong
author_facet Xiong, Wei
Ye, Chenlu
Liao, Baohao
Dong, Hanze
Xu, Xinxing
Monz, Christof
Bian, Jiang
Jiang, Nan
Zhang, Tong
contents Reinforcement learning (RL) for large language model reasoning is frequently hindered by signal loss, a phenomenon where standard uniform sampling with small group sizes fails to uncover informative learning signals for difficult prompts. We demonstrate that this collapse is a statistical artifact of undersampling rather than an inherent model limitation. To address this systematically, we introduce a theoretical framework based on optimizing a non-linear RL objective (e.g., log-likelihood). We show that this objective naturally induces a weighted gradient estimator that prioritizes difficult prompts, which can be robustly realized through adaptive sampling. Guided by this framework, we propose Reinforce-Ada, a family of algorithms that dynamically allocates inference budgets based on prompt difficulty, effectively scaling up RL compute to where it is needed most. Unlike passive filtering methods that discard low-signal prompts, Reinforce-Ada actively invests compute to recover them. We introduce two efficient realizations: an estimation-based approach and a model-free sequential sampling approach. Extensive experiments across multiple benchmarks show that Reinforce-Ada significantly outperforms uniform baselines like GRPO, recovering lost signals and accelerating convergence by up to $2\times$ while maintaining the same total inference budget. Code is available at https://github.com/RLHFlow/Reinforce-Ada.
format Preprint
id arxiv_https___arxiv_org_abs_2510_04996
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Reinforce-Ada: An Adaptive Sampling Framework under Non-linear RL Objectives
Xiong, Wei
Ye, Chenlu
Liao, Baohao
Dong, Hanze
Xu, Xinxing
Monz, Christof
Bian, Jiang
Jiang, Nan
Zhang, Tong
Machine Learning
Artificial Intelligence
Computation and Language
Reinforcement learning (RL) for large language model reasoning is frequently hindered by signal loss, a phenomenon where standard uniform sampling with small group sizes fails to uncover informative learning signals for difficult prompts. We demonstrate that this collapse is a statistical artifact of undersampling rather than an inherent model limitation. To address this systematically, we introduce a theoretical framework based on optimizing a non-linear RL objective (e.g., log-likelihood). We show that this objective naturally induces a weighted gradient estimator that prioritizes difficult prompts, which can be robustly realized through adaptive sampling. Guided by this framework, we propose Reinforce-Ada, a family of algorithms that dynamically allocates inference budgets based on prompt difficulty, effectively scaling up RL compute to where it is needed most. Unlike passive filtering methods that discard low-signal prompts, Reinforce-Ada actively invests compute to recover them. We introduce two efficient realizations: an estimation-based approach and a model-free sequential sampling approach. Extensive experiments across multiple benchmarks show that Reinforce-Ada significantly outperforms uniform baselines like GRPO, recovering lost signals and accelerating convergence by up to $2\times$ while maintaining the same total inference budget. Code is available at https://github.com/RLHFlow/Reinforce-Ada.
title Reinforce-Ada: An Adaptive Sampling Framework under Non-linear RL Objectives
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2510.04996