[ICLR 2025] SuperCorrect: Supervising and Correcting Language Models with Error-Driven Insights

Introduction

This repo provides the official implementation of SuperCorrect, a novel two-stage fine-tuning method for improving both reasoning accuracy and self-correction ability for LLMs. We incorporate LLMs with our pre-defined hierarchical thought template (Buffer of Thought (BoT)) to conduct more deliberate reasoning than conventional CoT.

📢 News

[2025.2] 🎉 We release ReasonFlux, a revolutionary inference-scaling paradigm with a hierarchical RL algorithm: enpowering a 32B model with 500 thought templates to outperform o1-preview and DeepSeek-V3 in reasoning tasks.

Task	ReasonFlux 32B	DeepSeek V3	OpenAI o1-preview	OpenAI o1-mini	QWQ 32B-preview	GPT 4o
MATH	91.2	90.2	85.5	90.0	90.6	76.6
AIME 2024	56.7	39.2	44.6	56.7	50.0	9.3
Olympiad Bench	63.3	55.4	-	65.3	61.2	43.3
GaokaoEn 2023	83.6	-	71.4	78.4	65.3	67.5
AMC2023	85.0	80.0	90.0	95.0	-	47.5

Quick Start

Installation

git clone https://github.com/YangLing0818/SuperCorrect
cd SuperCorrect
conda create -n SuperCorrect python==3.10
conda activate SuperCorrect
pip install -r requirements.txt

Requirements

• Since our current model is based on Qwen2.5-Math series, transformers>=4.37.0 is needed for Qwen2.5-Math models. The latest version is recommended.

Warning

🚨 This is a must because `transformers` integrated Qwen2 codes since `4.37.0`.

Inference with Different Library

🤗 Hugging Face Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "BitStarWalkin/SuperCorrect-7B"
device = "cuda" 

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "Find the distance between the foci of the ellipse \[9x^2 + \frac{y^2}{9} = 99.\]"
hierarchical_prompt = "Solve the following math problem in a step-by-step XML format, each step should be enclosed within tags like <Step1></Step1>. For each step enclosed within the tags, determine if this step is challenging and tricky, if so, add detailed explanation and analysis enclosed within <Key> </Key> in this step, as helpful annotations to help you thinking and remind yourself how to conduct reasoning correctly. After all the reasoning steps, summarize the common solution and reasoning steps to help you and your classmates who are not good at math generalize to similar problems within <Generalized></Generalized>. Finally present the final answer within <Answer> </Answer>."
# HT
messages = [
    {"role": "system", "content":hierarchical_prompt },
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=1024
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

🔥 vLLM

import os
from vllm import LLM, SamplingParams
model_name = 'BitStarWalkin/SuperCorrect-7B'
hierarchical_prompt = "Solve the following math problem in a step-by-step XML format, each step should be enclosed within tags like <Step1></Step1>. For each step enclosed within the tags, determine if this step is challenging and tricky, if so, add detailed explanation and analysis enclosed within <Key> </Key> in this step, as helpful annotations to help you thinking and remind yourself how to conduct reasoning correctly. After all the reasoning steps, summarize the common solution and reasoning steps to help you and your classmates who are not good at math generalize to similar problems within <Generalized></Generalized>. Finally present the final answer within <Answer> </Answer>."
prompts = [
    "For what positive value of $t$ is $|{-4+ti}| = 6$?",
    "Find the distance between the foci of the ellipse \\[9x^2 + \\frac{y^2}{9} = 99.\\]",
    "The fourth term of a geometric series is $24$ and the eleventh term is $3072$. What is the common ratio?"
]
combined_prompts = [hierarchial_prompt + '\n' + prompt for prompt in prompts]
sampling_params = SamplingParams(temperature=0, top_p=1,max_tokens=1024)
llm = LLM(model=model_name, trust_remote_code=True)
outputs = llm.generate(combined_prompts, sampling_params)

#Print the outputs.
for output in outputs:
    prompt = output.prompt
    generated_text = output.outputs[0].text
    print(f"Prompt: {prompt}")
    print(f"Generated text: {generated_text}")

Here we also provide inference code with vLLM . vLLM is a fast and easy-to-use library for LLM inference and serving.

1. Our evaluation

Here we provide two different evaluation methods: online version which utilizes GPT-4o to conduct a more fair and robust judgement and offline version which utilizes programming method to verify the final results. Both methods aim to provide a more accurate and strict evaluation results, as the final results in MATH dataset are not always numeric or pure expression. We now provide online version for evaluation, we will update soon for offline version.

API_KEY= "Input your key here"
MODEL_NAME_OR_PATH="BitStarWalkin/SuperCorrect-7B"
export CUDA_VISIBLE_DEVICES="0"
bash evaluation.sh $API_KEY $MODEL_NAME_OR_PATH

2. Evaluation with lm-evaluation-harness

lm_eval --model hf \
    --model_args pretrained="Qwen2.5-Math-7B-Instruct" \
    --tasks minerva_math \
    --log_samples \
    --output_path Qwen2.5-Math-7B-Instruct-lm-evaluation \
    --batch_size 12

lm_eval --model hf \
    --model_args pretrained="SuperCorrect-7B" \
    --tasks minerva_math \
    --log_samples \
    --output_path SuperCorrect-7B-lm-evaluation \
    --batch_size 12

Evaluation results produced by lm-evaluation:

Qwen2.5-Math-7B-Instruct	Version	Filter	n-shot	Metric		Value		Stderr
minerva_math	1	none	4	exact_match	↑	0.5034	±	0.0064
- minerva_math_algebra	1	none	4	exact_match	↑	0.7009	±	0.0133
- minerva_math_counting_and_prob	1	none	4	exact_match	↑	0.5232	±	0.0230
- minerva_math_geometry	1	none	4	exact_match	↑	0.4635	±	0.0228
- minerva_math_intermediate_algebra	1	none	4	exact_match	↑	0.2237	±	0.0139
- minerva_math_num_theory	1	none	4	exact_match	↑	0.4667	±	0.0215
- minerva_math_prealgebra	1	none	4	exact_match	↑	0.7394	±	0.0149
- minerva_math_precalc	1	none	4	exact_match	↑	0.2143	±	0.0176

SuperCorrect-7B	Version	Filter	n-shot	Metric		Value		Stderr
minerva_math	1	none	4	exact_match	↑	0.6188 (+0.1154)	±	0.0065
- minerva_math_algebra	1	none	4	exact_match	↑	0.7936 (+0.0927)	±	0.0118
- minerva_math_counting_and_prob	1	none	4	exact_match	↑	0.5802 (+0.0570)	±	0.0227
- minerva_math_geometry	1	none	4	exact_match	↑	0.5261 (+0.0626)	±	0.0228
- minerva_math_intermediate_algebra	1	none	4	exact_match	↑	0.4385 (+0.2148)	±	0.0165
- minerva_math_num_theory	1	none	4	exact_match	↑	0.6167 (+0.1500)	±	0.0209
- minerva_math_prealgebra	1	none	4	exact_match	↑	0.7715 (+0.0321)	±	0.0142
- minerva_math_precalc	1	none	4	exact_match	↑	0.4103 (+0.1960)	±	0.0211

Summary	Version	Filter	n-shot	Metric		Value		Stderr
Qwen2.5-Math-7B-Instruct	1	none	4	exact_match	↑	0.5034	±	0.0064
SuperCorrect-7B	1	none	4	exact_match	↑	0.6188 (+0.1154)	±	0.0065

3. Evaluation with Qwen2.5-Math-Evaluation

export CUDA_VISIBLE_DEVICES="0"
MODEL_NAME_OR_PATH="Qwen/Qwen2.5-Math-7B-Instruct"
bash sh/eval.sh $PROMPT_TYPE $MODEL_NAME_OR_PATH

export CUDA_VISIBLE_DEVICES="0"
MODEL_NAME_OR_PATH="BitStarWalkin/SuperCorrect-7B"
bash sh/eval.sh $PROMPT_TYPE $MODEL_NAME_OR_PATH

Evaluation results produced by Qwen2.5-Math-Eval:

Model	MATH Accuracy (%)
Qwen2.5-Math	80.6
SuperCorrect	82.1
Our Improvement	+1.5

Citation

@inproceedings{yang2025supercorrect,
  title={SuperCorrect: Supervising and Correcting Language Models with Error-Driven Insights},
  author={Yang, Ling and Yu, Zhaochen and Zhang, Tianjun and Xu, Minkai and Gonzalez, Joseph E and Cui, Bin and Yan, Shuicheng},
  booktitle={International Conference on Learning Representations},
  year={2025}
}

@article{yang2024buffer,
  title={Buffer of Thoughts: Thought-Augmented Reasoning with Large Language Models},
  author={Yang, Ling and Yu, Zhaochen and Zhang, Tianjun and Cao, Shiyi and Xu, Minkai and Zhang, Wentao and Gonzalez, Joseph E and Cui, Bin},
  journal={Advances in Neural Information Processing Systems},
  year={2024}
}

Acknowledgements

Our SuperCorrect is a two-stage fine-tuning model which based on several extraordinary open-source models like Qwen2.5-Math, DeepSeek-Math, Llama3-Series. Our evaluation method is based on the code base of outstanding works like Qwen2.5-Math and lm-evaluation-harness. We also want to express our gratitude for amazing works such as BoT which provides the idea of thought template.