Research map / Large language models / Language, safety and interpretability
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
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add bias-capture and disentanglement terms to compression-only objective (ablation) | DIBJudge · DIB objective ablation | bias severity S_bias: 0.124 -> 0.031 | Improved |
| add disentanglement term to compression-only objective (ablation) | DIBJudge · DIB objective ablation | accuracy: 85.25 -> 88.55 | Improved |
| add compression term to bias+disentangle (ablation) | DIBJudge · DIB objective ablation | accuracy: 89.10 -> 89.85 | Improved |
| add bias-capture term to compression+disentangle (ablation) | DIBJudge · DIB objective ablation | accuracy: 88.55 -> 89.85 | Improved |
| add disentanglement term to compression+bias (ablation) | DIBJudge · DIB objective ablation | bias severity S_bias: 0.091 -> 0.031 | Improved |
| add both proxy tasks CLA and LPBC (ablation) | DIBJudge · Proxy task ablation | bias severity S_bias: 0.421 -> 0.147 | Improved |
| add both proxy tasks CLA and LPBC (ablation) | DIBJudge · Proxy task ablation | accuracy: 87.12 -> 89.18 | Improved |
| add LPBC proxy task alone (ablation) | DIBJudge · Proxy task ablation | bias severity S_bias: 0.421 -> 0.279 | Improved |
| add CLA proxy task alone (ablation) | DIBJudge · Proxy task ablation | bias severity S_bias: 0.421 -> 0.312 | Improved |
| replace vanilla SFT with DIBJudge | Qwen3 judge · Skywork held-out, length bias | Spearman rho: 0.553 -> 0.314 | Improved |
| replace vanilla IB with DIBJudge | Qwen3 judge · Skywork held-out, self-preference | bias severity S_bias: 0.276 -> 0.219 | Improved |
| replace vanilla IB with DIBJudge | Qwen3 judge · Translationese bias | bias severity S_bias: 0.168 -> 0.083 | Improved |
Closest papers by meaning
- Biases Mitigation and Expressiveness Preservation in Language Models: A Comprehensive Pipeline (Student Abstract) (AAAI 2024 · similarity 0.87)
- UniBias: Unveiling and Mitigating LLM Bias through Internal Attention and FFN Manipulation (NeurIPS 2024 · similarity 0.87)
- Debiasing Algorithm through Model Adaptation (ICLR 2024 · similarity 0.87)
- Language Ranker: A Metric for Quantifying LLM Performance Across High and Low-Resource Languages (AAAI 2025 · similarity 0.87)
- Any Large Language Model Can Be a Reliable Judge: Debiasing with a Reasoning-based Bias Detector (NeurIPS 2025 · similarity 0.87)
- MuRating: A High Quality Data Selecting Approach to Multilingual Large Language Model Pretraining (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
