StepCo

We propose StepCo, a novel framework that uses an iterative verify-then-revise process to progressively identify and revise incorrect steps in LLM-generated reasoning paths.

Steps to Run the Code

1. Constructing Process Supervision Dataset

First, you need to switch to the Data_Synthesis directory. To generate the problem-solving steps, run the following command, which will use GPU 0 by default:

CUDA_VISIBLE_DEVICES=0 python run.py

After generating the problem-solving steps, run the following command to compute the probability of getting the correct answer for each step:

python tree_construction.py

This will output the probability associated with each reasoning step.

Note: You will need to specify the paths for reading and saving datasets, as well as the model name and configuration for inference steps.

2. Training Process Supervised Verifier

Then, you need to switch to the Process_Supervised_Verifier directory. Run the fine-tuning script using the following command:

CUDA_VISIBLE_DEVICES=0,1,2 python finetune_deberta_v3.py \
--model_name_or_path microsoft/deberta-v3-large \
--cache_dir xxx \
--output_dir xxx \
--shuffle_train_dataset \
--do_train \
--max_seq_length 256 \
--per_device_train_batch_size 16 \
--learning_rate 2e-6 \
--num_train_epochs 100 \
--pad_to_max_length \
--save_steps 500 \
--gradient_accumulation_steps 10

Note: Before running the code, specify the paths for saving the model and loading the dataset.

3. Solving Math Word Problems

Finally, you need to switch to the StepCo directory. Run the following command to automatically solve the math word problem:

CUDA_VISIBLE_DEVICES=2 python main.py --dataset_idx 0

Note: You need to specify the root directory where your test dataset is stored. For interacting with OpenAI's API, you must configure your API keys and any additional required settings.

@inproceedings{wu-etal-2025-stepco,
    title = "Enhancing Mathematical Reasoning in {LLM}s by Stepwise Correction",
    author = "Wu, Zhenyu  and
      Zeng, Qingkai  and
      Zhang, Zhihan  and
      Tan, Zhaoxuan  and
      Shen, Chao  and
      Jiang, Meng",
    editor = "Che, Wanxiang  and
      Nabende, Joyce  and
      Shutova, Ekaterina  and
      Pilehvar, Mohammad Taher",
    booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.acl-long.1048/",
    doi = "10.18653/v1/2025.acl-long.1048",
    pages = "21602--21623",
    ISBN = "979-8-89176-251-0",
}