Mitigating Translationese Bias in Multilingual LLM-as-a-Judge via Disentangled Information Bottleneck

Hongbin Zhang, Kehai Chen, Xuefeng Bai, Youcheng Pan, Yang Xiang, Jinpeng Wang, Min Zhang
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:154378-154416, 2026.

Abstract

Large language models (LLMs) have emerged as a standard paradigm for automated multilingual evaluation, yet exhibit systematic biases. In this paper, we identify “translationese bias”, in which LLMs systematically favor machine-translated text over human-authored references, and this bias is particularly pronounced in low-resource languages. We attribute this bias to spurious correlations with (a) strong latent manifold isomorphism with English and (b) high predictive confidence. To mitigate these issues, we present DIBJudge, a robust fine-tuning framework that decouples robust features and bias representations by explicitly isolating spurious attributes into a dedicated bias branch and penalizing mutual dependence to enforce disentanglement. In particular, we present a vector-quantized compression that ensures the robust representation retains minimal and sufficient judgment-critical information. Extensive evaluations on multilingual reward modeling benchmarks and a specially designed translationese bias evaluation suite demonstrate that DIBJudge outperforms strong baselines and effectively mitigates translationese bias.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-zhang26k, title = {Mitigating Translationese Bias in Multilingual {LLM}-as-a-Judge via Disentangled Information Bottleneck}, author = {Zhang, Hongbin and Chen, Kehai and Bai, Xuefeng and Pan, Youcheng and Xiang, Yang and Wang, Jinpeng and Zhang, Min}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {154378--154416}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/zhang26k/zhang26k.pdf}, url = {https://proceedings.mlr.press/v306/zhang26k.html}, abstract = {Large language models (LLMs) have emerged as a standard paradigm for automated multilingual evaluation, yet exhibit systematic biases. In this paper, we identify “translationese bias”, in which LLMs systematically favor machine-translated text over human-authored references, and this bias is particularly pronounced in low-resource languages. We attribute this bias to spurious correlations with (a) strong latent manifold isomorphism with English and (b) high predictive confidence. To mitigate these issues, we present DIBJudge, a robust fine-tuning framework that decouples robust features and bias representations by explicitly isolating spurious attributes into a dedicated bias branch and penalizing mutual dependence to enforce disentanglement. In particular, we present a vector-quantized compression that ensures the robust representation retains minimal and sufficient judgment-critical information. Extensive evaluations on multilingual reward modeling benchmarks and a specially designed translationese bias evaluation suite demonstrate that DIBJudge outperforms strong baselines and effectively mitigates translationese bias.} }
Endnote
%0 Conference Paper %T Mitigating Translationese Bias in Multilingual LLM-as-a-Judge via Disentangled Information Bottleneck %A Hongbin Zhang %A Kehai Chen %A Xuefeng Bai %A Youcheng Pan %A Yang Xiang %A Jinpeng Wang %A Min Zhang %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-zhang26k %I PMLR %P 154378--154416 %U https://proceedings.mlr.press/v306/zhang26k.html %V 306 %X Large language models (LLMs) have emerged as a standard paradigm for automated multilingual evaluation, yet exhibit systematic biases. In this paper, we identify “translationese bias”, in which LLMs systematically favor machine-translated text over human-authored references, and this bias is particularly pronounced in low-resource languages. We attribute this bias to spurious correlations with (a) strong latent manifold isomorphism with English and (b) high predictive confidence. To mitigate these issues, we present DIBJudge, a robust fine-tuning framework that decouples robust features and bias representations by explicitly isolating spurious attributes into a dedicated bias branch and penalizing mutual dependence to enforce disentanglement. In particular, we present a vector-quantized compression that ensures the robust representation retains minimal and sufficient judgment-critical information. Extensive evaluations on multilingual reward modeling benchmarks and a specially designed translationese bias evaluation suite demonstrate that DIBJudge outperforms strong baselines and effectively mitigates translationese bias.
APA
Zhang, H., Chen, K., Bai, X., Pan, Y., Xiang, Y., Wang, J. & Zhang, M.. (2026). Mitigating Translationese Bias in Multilingual LLM-as-a-Judge via Disentangled Information Bottleneck. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:154378-154416 Available from https://proceedings.mlr.press/v306/zhang26k.html.

Related Material