Official code for paper Revisiting Model Interpolation for Efficient Reasoning
Please follow the official guidance of Opencompass to set up a python environment.
We use the lmdeploy backend, please remember to set
pip install "opencompass[lmdeploy]"
Download the official weights from huggingface:
We recommend to download via the huggingface-cli, such as
hf download Qwen/Qwen3-30B-A3B-Thinking-2507 --token $your_hf_token --local-dir weights/Qwen3-30B-A3B/Qwen3-30B-A3B-Thinking-2507
hf download Qwen/Qwen3-30B-A3B-Instruct-2507 --token $your_hf_token --local-dir weights/Qwen3-30B-A3B/Qwen3-30B-A3B-Instruct-2507
Then, run the mi.py:
python mi.py --model_b /path/to/your/projects/base_model --model_i /path/to/your/projects/finetuned_model --lambda_val 0.5 --output_dir /path/to/your/projects/merged_output
where lambda_val is the interpolation factor.
We employ the opencompass for evaluation.
You need to modify the config files first.
For example, in evaluation/qwen3_AIME.py, replace the paths with your folder, modify the gpus to fit your machine.
Then all you need is to run opencompass evaluation/qwen3_AIME.py and wait the final results.
Warnning: In this repo, we benchmark the Instruct-2507/Thinking-2507 version, which do not require setting 'enable_thinking'. If you try to evaluate the Qwen3 hybrid thinking model, such as Qwen3-4B, please fix the bugs in Opencompass and pass an extra enable_thinking following this repo.
We use the Apache‑2.0 license. Please also comply with the licenses of any upstream models and datasets.
If you find this repository helpful, please consider citing our paper:
@article{wu2025revisiting,
title={Revisiting Model Interpolation for Efficient Reasoning},
author={Wu, Taiqiang and Yang, Runming and Liu, Tao and Wang, Jiahao and Wong, Ngai.},
journal={arXiv preprint arXiv:2510.10977},
year={2025}
}
For any questions, please pull an issue or email at takiwu@connect.hku.hk
