63 lines
2.4 KiB
Markdown
63 lines
2.4 KiB
Markdown
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
tags:
|
||
|
|
- large-language-model
|
||
|
|
- multi-agent-systems
|
||
|
|
- reinforcement-learning
|
||
|
|
- agentic-ai
|
||
|
|
- code
|
||
|
|
- math
|
||
|
|
---
|
||
|
|
|
||
|
|
# MetaAgent-X: Breaking the Ceiling of Automatic Multi-Agent Systems via End-to-End Reinforcement Learning
|
||
|
|
|
||
|
|
[Paper 📑](https://arxiv.org/abs/2605.14212)
|
||
|
|
|
||
|
|
[Codebase 🚗](https://github.com/pettingllms-ai/PettingLLMs)
|
||
|
|
|
||
|
|
[Project Page 🏆](https://mercury7353.github.io/MetaAgent-X-Page/)
|
||
|
|
|
||
|
|
## Overview
|
||
|
|
|
||
|
|
**MetaAgent-X** is an end-to-end reinforcement learning framework for autonomous multi-agent systems.
|
||
|
|
|
||
|
|
Unlike conventional automatic MAS methods that rely on frozen models, hand-crafted prompts, or search-based workflows, MetaAgent-X trains one shared model to both **design** a multi-agent system and **execute** it. The model learns to generate task-adaptive agent roles, collaboration structures, and execution strategies through reinforcement learning.
|
||
|
|
|
||
|
|
MetaAgent-X demonstrates strong cross-domain adaptation and achieves state-of-the-art performance across both **code** and **math** benchmarks.
|
||
|
|
|
||
|
|
## Key Features
|
||
|
|
|
||
|
|
- **One model for both design and execution**: the same model acts as both the MAS designer and the task executor.
|
||
|
|
- **End-to-end reinforcement learning**: the model is optimized directly from downstream task outcomes.
|
||
|
|
- **Autonomous multi-agent system generation**: the model learns to construct and execute agent swarms for complex reasoning tasks.
|
||
|
|
- **Cross-domain generalization**: strong performance on both coding and mathematical reasoning benchmarks.
|
||
|
|
|
||
|
|
## Results
|
||
|
|
|
||
|
|
The following table reports the performance of **MetaAgent-X<sub>RL</sub>**.
|
||
|
|
Numbers in parentheses denote absolute gains over the single-agent baseline.
|
||
|
|
|
||
|
|
| Domain | Benchmark | MetaAgent-X<sub>RL</sub> |
|
||
|
|
|---|---:|---:|
|
||
|
|
| Code | LiveCodeBench | **41.00** |
|
||
|
|
| Code | APPS | **38.00** |
|
||
|
|
| Code | CodeContests | **17.00** |
|
||
|
|
| Math | AIME24 | **40.00** |
|
||
|
|
| Math | AIME25 | **33.33** |
|
||
|
|
| Math | OlympiadBench | **61.00** |
|
||
|
|
| Overall | Average | **38.33** |
|
||
|
|
|
||
|
|
## Citation
|
||
|
|
```
|
||
|
|
@misc{zhang2026metaagentxbreakingceiling,
|
||
|
|
title={MetaAgent-X : Breaking the Ceiling of Automatic Multi-Agent Systems via End-to-End Reinforcement Learning},
|
||
|
|
author={Yaolun Zhang and Yujie Zhao and Nan Wang and Yiran Wu and Jiayu Chang and Yizhao Chen and Qingyun Wu and Jishen Zhao and Huazheng Wang},
|
||
|
|
year={2026},
|
||
|
|
eprint={2605.14212},
|
||
|
|
archivePrefix={arXiv},
|
||
|
|
primaryClass={cs.AI},
|
||
|
|
url={https://arxiv.org/abs/2605.14212},
|
||
|
|
}
|
||
|
|
```
|