Self-Training Elicits Concise Reasoning in Large Language Models

This model is fine-tuned using self-training methods to generate concise reasoning paths for reasoning tasks while maintaining accuracy.

Model Details

Developed by: Tergel Munkhbat, Namgyu Ho, Seo Hyun Kim, Yongjin Yang, Yujin Kim, Se-Young Yun at KAIST AI
Model type: Fine-tuned Large Language Model for concise reasoning
Language(s) (NLP): English
License: MIT
Finetuned from model: Qwen/Qwen2.5-3B-Instruct
Repository: https://github.com/TergelMunkhbat/concise-reasoning
Paper: Self-Training Elicits Concise Reasoning in Large Language Models

How to Get Started with the Model

Use the code below to get started with the model.

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_name = "tergel/qwen2.5-3b-instruct-math-fs-gpt4o-bon"
device = "cuda" if torch.cuda.is_available() else "cpu"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map=device, torch_dtype=torch.bfloat16)

question = "Convert the point $(0,3)$ in rectangular coordinates to polar coordinates. Enter your answer in the form $(r,\\theta),$ where $r > 0$ and $0 \\le \\theta < 2 \\pi.$"

inputs = tokenizer(question, return_tensors="pt").to(device)
input_length = len(inputs['input_ids'][0])

outputs = model.generate(**inputs, max_new_tokens=512)

response = tokenizer.decode(outputs[0][input_length:], skip_special_tokens=True)
print(response)

For more detailed information about training methods, evaluation results, limitations, and technical specifications, please refer to our paper.

Citation

@article{munkhbat2025self,
  title={Self-Training Elicits Concise Reasoning in Large Language Models},
  author={Munkhbat, Tergel and Ho, Namgyu and Kim, Seohyun and Yang, Yongjin and Kim, Yujin and Yun, Se-Young},
  journal={arXiv preprint arXiv:2502.20122},
  year={2025}
}