deokhk/mmbert_ft_understandability_Qwen3-4B

mmBERT Detector for "Why Do Multilingual Reasoning Gaps Emerge in Reasoning Language Models?"

This repository provides mmBERT-based detectors introduced in our paper "Why Do Multilingual Reasoning Gaps Emerge in Reasoning Language Models?".
We release mmBERT detector checkpoints trained for Qwen3-4B, using two training setups:

• mgsm_filtered
• mmlu_prox_lite_dev

For each setup, we provide three independent seeds:

• seed 32
• seed 42
• seed 52

These detectors are intended for research use in analyzing multilingual reasoning gaps and understandability-related behaviors.

Citation

If you find this repository useful, please cite:

@misc{kang2025multilingualreasoninggapsemerge,
      title={Why Do Multilingual Reasoning Gaps Emerge in Reasoning Language Models?}, 
      author={Deokhyung Kang and Seonjeong Hwang and Daehui Kim and Daehui Kim and Hyounghun Kim and Gary Geunbae Lee},
      year={2025},
      eprint={2510.27269},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.27269}, 
}