ExGRPO-LUFFY-7B-Continual

ExGRPO: Learning to Reason from Experience

This repository contains the model based on the ExGRPO framework, presented in the paper ExGRPO: Learning to Reason from Experience. ExGRPO (Experiential Group Relative Policy Optimization) is a novel paradigm for improving the reasoning ability of large language models (LLMs) by addressing the computational inefficiency and instability of standard on-policy training in Reinforcement Learning from Verifiable Rewards (RLVR).

ExGRPO strategically manages and replays high-value experiences through bucketed prioritization and mixed-policy optimization, enabling more efficient and stable RLVR training.

Code: https://github.com/ElliottYan/LUFFY/tree/main/ExGRPO

<div align="center"> <img src="https://github.com/ElliottYan/LUFFY/raw/main/ExGRPO/figures/exgrpo_intro.png" alt="ExGRPO Overview" style="width: 88%; height: auto;"> </div>

---

Introduction

Existing RLVR methods for reasoning tasks predominantly rely on on-policy optimization, which discards online rollouts after a single update, wasting valuable exploration signals and constraining scalability. We conduct a systematic analysis of experience utility in RLVR and identify question difficulty and trajectory entropy as effective online proxies for assessing experience quality. Building on these insights, we propose ExGRPO, a novel framework that strategically manages and replays high-value experiences through bucketed prioritization and mixed-policy optimization, enabling more efficient and stable RLVR training.

Key Highlights:

• Experience Value Modeling: Introduces the online proxy metrics: rollout correctness and trajectory entropy, for quantifying the value of RLVR experience.
• ExGRPO Framework: Built on top of GRPO, ExGRPO introduces a systematic experience management mechanism and an experience optimization objective to maximize the benefit of past explorations.
• Generalization and Stability: Demonstrates broad applicability across different backbone models and mitigates training collapse of on-policy RLVR in challenging scenarios.

---

Getting Started

Installation

You can install dependencies by running the following commands:

conda create -n exgrpo python=3.10
conda activate exgrpo
cd exgrpo
pip install -r requirements.txt
pip install -e .
cd verl
pip install -e .

Note: If you encounter issues caused by the pyairports library, please refer to this hot-fix solution.

For the flash-attn library, we use the v2.7.4-post1 release and recommend installing it via the pre-built wheel. Please adjust based on your environment.

wget https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
pip install flash_attn-2.7.4.post1+cu12torch2.4cxx11abiFALSE-cp310-cp310-linux_x86_64.whl

Usage

Data Preparation

You need to first run the data preparation script to get the training data in parquet format.

cd data
python prepare_train.py --dataset_name Elliott/Openr1-Math-46k-8192 --output_file openr1.parquet

Note: Although we utilize the OpenR1 data, only the question field is used in RLVR. The ExGRPO data processing pipeline does not incorporate the external R1 trajectory during training.

---

Evaluation

Main Results

Zero RLVR on Qwen2.5-Math-7B & Continual RLVR on LUFFY

<div align="center"> <img src="https://github.com/ElliottYan/LUFFY/raw/main/ExGRPO/figures/main_result.png" alt="Main Results" style="width: 95%; height: auto;"> </div>

Zero RLVR on Llama3.1-8B (Base, Instruct), Qwen2.5-Math 1.5B Base, Qwen2.5-7B Instruct

<div align="center"> <img src="https://github.com/ElliottYan/LUFFY/raw/main/ExGRPO/figures/model_extensions_bar.png" alt="Model Extensions Bar" style="width: 95%; height: auto;"> </div> <details> <summary>Click to view full results of model extension</summary> <div align="center"> <img src="https://github.com/ElliottYan/LUFFY/raw/main/ExGRPO/figures/model_extensions.png" alt="Full Model Extensions" style="width: 95%; height: auto;"> </div> </details>

Released Models

---

Citation

If you find our model, data, or evaluation code useful, please kindly cite our paper:

@article{zhan2025exgrpo,
      title={ExGRPO: Learning to Reason from Experience}, 
      author={Runzhe Zhan and Yafu Li and Zhi Wang and Xiaoye Qu and Dongrui Liu and Jing Shao and Derek F. Wong and Yu Cheng},
      year={2025},
      journal = {ArXiv preprint},
      volume = {2510.02245},
      url={https://arxiv.org/abs/2510.02245}, 
}