SII-Enigma/Llama3.2-8B-Ins-AMPO

Introduction

AMPO, a novel framework that intelligently leverages guidance from multiple, diverse teacher models, intervening only when the on-policy model fails. Our two core contributions, Adaptive Multi-Guidance Replacement and Comprehension-based Guidance Selection, ensure that this external knowledge is used both efficiently and effectively.

Key Highlights:

• Adaptive Multi-Guidance Replacement: Minimizes intervention by providing external guidance only upon complete on-policy failure, preserving self-discovery while enhancing reasoning efficiency.
• Comprehension-based Guidance Selection: Improves learning effectiveness by guiding the model to assimilate the most comprehensible external solutions, demonstrably boosting performance.
• Superior Performance: Achieves better performance and efficiency compared to using RL or SFT alone.

Multi-Guidance Pool

Teacher Models: AceReason-Nemotron-1.1-7B, DeepSeek-R1-Distill-Qwen-7B, OpenR1-Qwen-7B, Qwen3-8B(thinking)

Inference Example

Here’s an example of using AMPO for inference:

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

model_path = "SII-Enigma/Llama3.2-8B-Ins-AMPO"

question = "which number is larger? 9.11 or 9.9?"

tokenizer = AutoTokenizer.from_pretrained(model_path)
messages = [{"role": "user", "content": question}]
chat = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

llm = LLM(model=model_path)
params = SamplingParams(temperature=0.6, max_tokens=8192)
outputs = llm.generate([chat], params)
print(outputs[0].outputs[0].text)

Acknowledgement

AMPO builds upon LUFFY, veRL, RLPR and utilizes vLLM for inference. We utilize Math-Verify for math reasoning evaluation. We thank the open-source community for codes, datasets and backbones.

Citation

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

@misc{yuan2025teacheradaptivemultiguidancepolicy,
      title={More Than One Teacher: Adaptive Multi-Guidance Policy Optimization for Diverse Exploration}, 
      author={Xiaoyang Yuan and Yujuan Ding and Yi Bin and Wenqi Shao and Jinyu Cai and Jingkuan Song and Yang Yang and Heng Tao Shen},
      year={2025},
      eprint={2510.02227},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.02227}, 
}