m1-32b

Two Heads are Better Than One: Test-time Scaling of Multi-agent Collaborative Reasoning

M1-32B is a 32B-parameter large language model fine-tuned from Qwen2.5-32B-Instruct on the M500 dataset—an interdisciplinary multi-agent collaborative reasoning dataset. M1-32B is optimized for improved reasoning, discussion, and decision-making in multi-agent systems (MAS), including frameworks such as AgentVerse.

Code: https://github.com/jincan333/MAS-TTS Project page: https://github.com/jincan333/MAS-TTS

---

How to Use with 🤗 Transformers

You can use this model directly with the transformers library for text generation.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Can111/m1-32b"

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16, # Use bfloat16 for optimal performance if supported
    device_map="auto" # Automatically distribute model across available devices
)
model.eval() # Set model to evaluation mode

# Define your conversation messages
messages = [
    {"role": "user", "content": "Explain multi-agent collaborative reasoning and its benefits."},
]

# Apply chat template and tokenize inputs
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# Generate response
generated_ids = model.generate(
    model_inputs.input_ids,
    max_new_tokens=256,
    do_sample=True,
    temperature=0.7,
    top_p=0.9
)

# Decode and print the generated text
decoded_output = tokenizer.batch_decode(generated_ids[:, model_inputs.input_ids.shape[1]:], skip_special_tokens=True)[0]
print(decoded_output)

---

🚀 Key Features

• 🧠 Enhanced Collaborative Reasoning
  Trained on real multi-agent traces involving diverse roles like Expert Recruiter, Problem Solvers, and Evaluator.

• 🗣️ Role-Aware Dialogue Generation
  Learns to reason and respond from different expert perspectives based on structured prompts.

• ⚙️ Optimized for Multi-Agent Systems
  Performs well as a MAS agent with adaptive collaboration and token budgeting.

---

🏗️ Model Training

• Base Model: Qwen2.5-32B-Instruct
• Dataset: M500 (500 curated multi-agent reasoning traces)
• Objective: Supervised Fine-Tuning (SFT) on role-conditioned prompts
• Training Setup:
  • 8 × A100 GPUs
  • 5 epochs
  • Learning rate: 1e-5
  • Frameworks: DeepSpeed, FlashAttention, LLaMA-Factory

---

📊 Performance

Table Caption:
Performance comparison on general understanding, mathematical reasoning, and coding tasks using strong reasoning and non-reasoning models within the AgentVerse framework. Our method achieves substantial improvements over Qwen2.5 and s1.1-32B on all tasks, and attains performance comparable to o3-mini and DeepSeek-R1 on MATH-500 and MBPP-S, demonstrating its effectiveness in enhancing collaborative reasoning in MAS. Note that the results of s1.1-32B are obtained without using budget forcing.

---

💬 Intended Use

M1-32B is intended for research on Multi-agent reasoning and collaboration in MAS

---

Citation

If you use this model, please cite the relevant papers:

@article{jin2025two,
  title={Two Heads are Better Than One: Test-time Scaling of Multi-agent Collaborative Reasoning},
  author={Jin, Can and Peng, Hongwu and Zhang, Qixin and Tang, Yujin and Metaxas, Dimitris N and Che, Tong},
  journal={arXiv preprint arXiv:2504.09772},
  year={2025}
}