AAPO

KO | EN

gitlite — search

#javascript #react #typescript #llm #deep-learning #linux #docker #cli #kubernetes #vue #pytorch #ios #mcp #open-source

(★ 16)

AAPO: Enhancing the Reasoning Capabilities of LLMs with Advantage Margin (ACL 2026)

AAPO 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# AAPO: Enhancing the Reasoning Capabilities of LLMs with Advantage Margin

**Advantage-Augmented Policy Optimization (AAPO)** — RL training algorithm that optimizes cross-entropy loss using advantages refined by a margin-based estimation scheme, reducing inefficiencies from group relative advantage estimation.

---

## Paper

| | |
| --- | --- |
| **Title** | AAPO: Enhancing the Reasoning Capabilities of LLMs with Advantage Margin |
| **Venue** | ACL 2026 Main Conference|
| **ArXiv** | [2505.14264](https://arxiv.org/abs/2505.14264) |
| **Authors** | $\text{Jian Xiong}$, $\text{Jingbo Zhou}^{\dagger}$, $\text{Jingyong Ye}$, $\text{Qiang Huang}$, $\text{Dejing Dou}^{\dagger}$ |
| **Models** | [Huggingface](https://huggingface.co/collections/jianxxiong/aapo) |

---

## Introduction

We propose **AAPO**, a novel RL algorithm for enhancing LLMs' reasoning capabilities. Experiments across multiple benchmarks and model families show consistent gains over strong baselines.

### Main results

Benchmark performance and statistical analysis of the zero advantage proportion during training:

![Main benchmark results](https://raw.githubusercontent.com/JianxXiong/AAPO/master/images/1.png)

![Zero-advantage proportion during training](https://raw.githubusercontent.com/JianxXiong/AAPO/master/images/2.png)

For training stability, convergence, ablations, OOD behavior, and training dynamics, please see the full paper.

---

## Repository layout

| Directory | Typical use |
| --- | --- |
| [`open-rs/`](open-rs) | Train **DeepSeek-R1-Distill-Qwen-1.5B** |
| [`open-r1/`](open-r1) | Train **Qwen2.5-Math-7B**, **Llama** (with config tweaks below), evaluation scripts |

Shared dependency list: [`requirements.txt`](requirements.txt).

---

## Quick start

### 1. Environment

```bash
conda create -n train python=3.11
conda activate train
pip install -r requirements.txt
```

### 2. Training

**DeepSeek-R1-Distill-Qwen-1.5B**

```bash
cd open-rs
bash train.sh
```

**Qwen2.5-Math-7B**

```bash
cd open-r1
bash train.sh
```

**Llama series**

In `train.sh`, set `max_completion_length=3072` and `max_prompt_length=1024`. In the config, set `dataset_name` to `SimpleRL-Zoo-Data/simplelr_abel_level3to5` and clear the system prompt. Then:

```bash
cd open-r1
bash train.sh
```

### 3. Evaluation

**Qwen series** — run inside the same tree you used for training (`open-r1` for Qwen2.5-Math-7B / Llama configs, `open-rs` for DeepSeek-R1-Distill-Qwen-1.5B).

Single benchmark:

```bash
cd open-r1   # or: cd open-rs
bash single_eval.sh
```

All benchmarks:

```bash
cd open-r1   # or: cd open-rs
bash auto_eval.sh
```

**Llama series**

Follow the evaluation setup in [SimpleRL-Reason](https://github.com/hkust-nlp/simpleRL-reason).

---

## Citation

If you find this work useful, please cite:

```bibtex
@inproceedings{
xiong2026aapoenhancingreasoningcapabilities,
title={AAPO: Enhancing the Reasoning Capabilities of LLMs with Advantage Margin},
author={Xiong, Jian and Zhou, Jingbo and Ye, Jingyong and Huang, Qiang and Dou, Dejing},
booktitle={The 64th Annual Meeting of the Association for Computational Linguistics},
year={2026}
}
```

---

## Acknowledgments

We thank the authors of [SimpleRL-Reason](https://github.com/hkust-nlp/simpleRL-reason), [open-rs](https://github.com/knoveleng/open-rs), and [GPG](https://github.com/AMAP-ML/GPG) for foundational code. Use of their resources is subject to their respective licenses.

---

## Contact

If you have any questions, please contact: jianxiong_ AT outlook DOT com.

AAPO

AAPO 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

AAPO

AAPO 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README