mlsanityData to 7 Oct 2026

Lower-Resource, Higher Scores: Language Bias in LLM Evaluators

arXiv 2607.14480 ↗First posted 16 Jul 2026cs.CLcited by 0

On the map: LLM alignment / RLHF / preference optimisation → Human values and AI feedback alignment

Abstract (arXiv)

LLM evaluators (trained reward models and prompted LLM-as-a-Judge) are routinely validated via pairwise accuracy. In a multilingual setting, this operates under the premise that high pairwise accuracy implies reliable, language-neutral scoring. We show that this assumption does not hold. We conduct experiments with semantically identical instruction-response pairs across 23 languages, and find that multilingual evaluators assign significantly different scores to different evaluation languages. The bias is statistically significant and consistent across eight open-weight evaluators of different architectures and training paradigms, persists in frontier judges, and is strongly correlated with language resource level: lower-resource languages are scored more generously. Meanwhile, these biases are invisible to pairwise accuracy: evaluators achieve above 90% pairwise accuracy, yet have up to 43% difference in acceptance rate across languages under a global decision threshold, meaning, for instance, that harmful content in lower-resource languages is more likely to pass safety filters. Per-language thresholds would require language identification, which can be defeated by code-switched prompts. We then investigate why lower-resource languages receive higher rather than lower scores, and we find that model uncertainty is linked with the effect: models tend to give higher scores when less confident, both under negative log-likelihood and under token-free uncertainty measures; however, language identity remains a significant predictor after controlling for uncertainty, and the bias cannot be explained away by content difficulty alone, but is a structural, language-level misalignment.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Human values and AI feedback alignment, one of 10 directions in LLM alignment / RLHF / preference optimisation. Builds on 43 of its 53 references that are in the map. First posted on arXiv 16 Jul 2026.

Builds on

43 of 53 references.

Closest ancestors · 2

Same direction, newest first

  1. 2026Mitigating Cross-Lingual Cultural Inconsistencies in LLMs via Consensus-Driven Preference Optimisation2 Apr
  2. 2023Prometheus: Inducing Fine-grained Evaluation Capability in Language Models12 Oct

Other references · 41

  1. 2022Training language models to follow instructions with human feedbackcited by 15,712
  2. 2023Judging LLM-as-a-Judge with MT-Bench and Chatbot Arenacited by 8,393
  3. 2015Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learningcited by 4,416
  4. 2016Simple and Scalable Predictive Uncertainty Estimation using Deep Ensemblescited by 3,888
  5. 2023Large Language Models are not Fair Evaluatorscited by 1,050
  6. 2022Red Teaming Language Models with Language Modelscited by 983
  7. 2023Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generationcited by 795
  8. 2020The State and Fate of Linguistic Diversity and Inclusion in the NLP Worldcited by 789
  9. 2023Safe RLHF: Safe Reinforcement Learning from Human Feedbackcited by 630
  10. 2023Language Is Not All You Need: Aligning Perception with Language Modelscited by 543
  11. 2022Improving alignment of dialogue agents via targeted human judgementscited by 527
  12. 2024RewardBench: Evaluating Reward Models for Language Modelingcited by 480
29 more
  1. 2024From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judgecited by 480
  2. 2023XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Modelscited by 477
  3. 2023Large Language Models Are State-of-the-Art Evaluators of Translation Qualitycited by 416
  4. 2023JudgeLM: Fine-tuned Large Language Models are Scalable Judgescited by 351
  5. 2024Skywork-Reward: Bag of Tricks for Reward Modeling in LLMscited by 322
  6. 2020How Good is Your Tokenizer? On the Monolingual Performance of Multilingual Language Modelscited by 313
  7. 2023MEGA: Multilingual Evaluation of Generative AIcited by 291
  8. 2023Evaluating Large Language Models at Evaluating Instruction Followingcited by 290
  9. 2023Language Model Tokenizers Introduce Unfairness Between Languagescited by 207
  10. 2023Do All Languages Cost the Same? Tokenization in the Era of Commercial Language Modelscited by 186
  11. 2022A New Generation of Perspective API: Efficient Multilingual Character-level Transformerscited by 181
  12. 2023OctoPack: Instruction Tuning Code Large Language Modelscited by 178
  13. 2024Large Language Models are Inconsistent and Biased Evaluatorscited by 146
  14. 2024Aya Expanse: Combining Research Breakthroughs for a New Multilingual Frontiercited by 121
  15. 2021Systematic Inequalities in Language Technology Performance across the World's Languagescited by 118
  16. 2024OffsetBias: Leveraging Debiased Data for Tuning Evaluatorscited by 94
  17. 2024Uncertainty-aware Reward Model: Teaching Reward Models to Know What is Unknowncited by 71
  18. 2023A Survey of Code-switching: Linguistic and Social Perspectives for Language Technologiescited by 59
  19. 2024M-RewardBench: Evaluating Reward Models in Multilingual Settingscited by 50
  20. 2025M-Prometheus: A Suite of Open Multilingual LLM Judgescited by 37
  21. 2025How Reliable is Multilingual LLM-as-a-Judge?cited by 29
  22. 2024ToxiCraft: A Novel Framework for Synthetic Generation of Harmful Informationcited by 17
  23. 2025Beyond the Final Layer: Intermediate Representations for Better Multilingual Calibration in Large Language Modelscited by 9
  24. 2025TRIDENT: Benchmarking LLM Safety in Finance, Medicine, and Lawcited by 7
  25. 2025Privacy-R1: Privacy-Aware Multi-LLM Agent Collaboration via Reinforcement Learningcited by 6
  26. 2024ToxiLab: How Well Do Open-Source LLMs Generate Synthetic Toxicity Data?cited by 5
  27. 2026CommonLID: Re-evaluating State-of-the-Art Language Identification Performance on Web Datacited by 5
  28. 2026Artificial intelligence is creating a new global linguistic hierarchycited by 3
  29. 2025Bias Beyond English: Evaluating Social Bias and Debiasing Methods in a Low-Resource Settingcited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2024LLM-as-a-Judge & Reward Model: What They Can and Cannot Do
  2. 2025Evaluating Scoring Bias in LLM-as-a-Judge
  3. 2024MM-Eval: A Multilingual Meta-Evaluation Benchmark for LLM-as-a-Judge and Reward Models
  4. 2026When Languages Disagree: Self-Evolving Multilingual LLM Judges
  5. 2025Do LLM Evaluators Prefer Themselves for a Reason?
  6. 2025Pairwise or Pointwise? Evaluating Feedback Protocols for Bias in LLM-Based Evaluation
  7. 2024Self-Preference Bias in LLM-as-a-Judge
  8. 2026Exploring the Effects of Alignment on Numerical Bias in Large Language Models