|
|
--- |
|
|
license: mit |
|
|
base_model: |
|
|
- Qwen/Qwen3-30B-A3B-Thinking-2507 |
|
|
--- |
|
|
<div align="center"> |
|
|
|
|
|
# MarsRL |
|
|
<div> |
|
|
Advancing <strong>M</strong>ulti-<strong>A</strong>gent <strong>R</strong>easoning <strong>S</strong>ystem via <strong>R</strong>einforcement <strong>L</strong>earning with Agentic Pipeline Parallelism |
|
|
</div> |
|
|
<a href="https://arxiv.org/pdf/2511.11373" target="_blank">Paper</a> | <a href="https://github.com/liushulinle/MarsRL" target="_blank">GitHub</a> |
|
|
</div> |
|
|
|
|
|
## Overview |
|
|
<hr /> |
|
|
Recent progress in large language models (LLMs) has been propelled by reinforcement learning with verifiable rewards (RLVR) and test-time scaling. However, the limited output length of LLMs constrains the depth of reasoning attainable in a single inference process. Multi-agent reasoning systems offer a promising alternative by employing multiple agents including Solver, Verifier, and Corrector, to iteratively refine solutions. While effective in closed-source models like Gemini 2.5 Pro, they struggle to generalize to open-source models due to insufficient critic and correction capabilities. To address this, we propose MarsRL, a novel reinforcement learning framework with agentic pipeline parallelism, designed to jointly optimize all agents in the system. MarsRL introduces agent-specific reward mechanisms to mitigate reward noise and employs pipeline-inspired training to enhance efficiency in handling long trajectories. Applied to Qwen3-30B-A3B-Thinking-2507, MarsRL improves AIME2025 accuracy from 86.5\% to 93.3\% and BeyondAIME from 64.9\% to 73.8\%, even surpassing Qwen3-235B-A22B-Thinking-2507. These findings highlight the potential of MarsRL to advance multi-agent reasoning systems and broaden their applicability across diverse reasoning tasks. |
|
|
<div align="center"> |
|
|
<img src="home.jpg" width="80%" /> |
|
|
</div> |
|
|
|
|
|
## V-C Reasoning System Evaluation Instructions |
|
|
<hr /> |
|
|
|
|
|
### step1: Download our released model or other open source models |
|
|
Supported models: Qwen3/DeepSeekV3.1/DeepSeek R1. You can modify the llm_client.py to use other models. |
|
|
|
|
|
### step2: Deploy service via VLLM |
|
|
|
|
|
### step3: Run the V-C reasoning system by the following commands: |
|
|
``` |
|
|
python3 vc_reasoning_system.py solver_ip_port_1,solver_ip_port_2,... vc_ip_port_1,vc_ip_port_2,... test_file output_dir |
|
|
for example: python3 vc_reasoning_system.py 8.8.8.8:8021,12.34.56.78:8021 8.8.8.8:8021,12.34.56.78:8021 ./outputs/debug ./test_corpus/aime2025.jsonl |
|
|
``` |
|
|
This step will run the reasoning system for each problem in the given $test_file$, the predicted results can be found in the output_dir |
|
|
|
|
|
### step4: Extract final solutions by the following commands: |
|
|
``` |
|
|
python3 extract_solution.py result_dir test_file |
|
|
for example: python3 extract_solution.py ./outputs/debug ./test_corpus/aime_2025.jsonl |
|
|
``` |
|
|
This step will generate a file named "eval_overalljsonl" in the input_dir. Your can evaluate the metrics based on this file. |
|
|
|
|
|
## Acknowledgements |
|
|
<hr /> |
|
|
|
|
|
- Our implementation is heaviliy built on [verl](https://github.com/volcengine/verl). |
|
|
|
|
|
- Our models are trained on top of [Qwen3-30B-A3B-Thinking-2507](https://huggingface.co/Qwen/Qwen3-30B-A3B-Thinking-2507). |
|
|
|
|
|
- Our V-C Reasoning system is built on [IMO25 pipline](https://github.com/lyang36/IMO25). |
|
|
|
|
|
Thanks for their wonderful work. |
|
|
|
|
|
## Citation |
|
|
<hr /> |
|
|
|
|
|
```bibtex |
|
|
@article{Marsrl2025, |
|
|
title = {MarsRL: Advancing Multi-Agent Reasoning System via Reinforcement Learning with Agentic Pipeline Parallelism}, |
|
|
author = {Shulin Liu, Dong Du, Tao Yang, Yang Li, Boyu Qiu} |
|
|
year = {2025} |
|
|
} |
|
|
``` |