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

• URL: http://arxiv.org/abs/2607.14480v2
• Date: Tue, 21 Jul 2026 06:28:06 GMT
• Title: Lower-Resource, Higher Scores: Language Bias in LLM Evaluators
• Abstract summary: We conduct experiments with semantically identical instruction-response pairs across 23 languages.<n>We find that multilingual evaluators assign significantly different scores to different evaluation languages.
• Score: 27.959960538188753
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: 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.

Related papers

• Choosing the Right Language Mode at Inference Time for Multilingual Reliability [3.6095388702618405]
  Large language models often struggle to reason in low- to mid-resource languages.<n>Prior work has shown that translation can improve multilingual reasoning by helping models access stronger English-centric representations.<n>This raises a central question: How much translation is needed for multilingual large language models to reason reliably.
  arXiv  Detail & Related papers  (2026-09-04T02:36:26Z)
• M-GATE: Multilingual Grammar, Accuracy in Translation, and Efficiency Benchmark for Large Language Models [0.0]
  M-GATE is a benchmark of linguistic proficiency spanning 30 typologically diverse languages.<n>We evaluate over 50 models in more than 80 configurations.
  arXiv  Detail & Related papers  (2026-08-04T15:16:05Z)
• Inference-Time Steering for Cross-Lingual Factual Consistency in LLMs [51.56484100374058]
  Large Language Models (LLMs) demonstrate remarkable multilingual fluency, but their internal knowledge representations remain disproportionately biased toward high-resource languages.<n>We investigate whether these biases can be mitigated at inference time, forcing an English-prompted model to answer as if it were queried in target languages.<n>Experiments on Gemma 3 12B Instruct reveal persona prompting to be the strongest overall intervention, balancing efficacy, safety, and out-of-domain generalization.
  arXiv  Detail & Related papers  (2026-07-21T16:15:05Z)
• Extending Item Response Theory for Efficient and Meaningful Multilingual Evaluation [19.965981631741894]
  Multilingual benchmarks are central to evaluating large language models (LLMs) across languages.<n>They suffer from three issues: exhaustive evaluation scales linearly with the number of languages, automatic translation introduces errors that are easily missed at scale, and some items conflate general and culture-specific knowledge.<n>We address all three with a unified statistical framework, Multilingual-IRT, which extends Item Response Theory with per-language difficulty deviations.
  arXiv  Detail & Related papers  (2026-06-14T07:16:03Z)
• Shared Doubt: Zero-shot Cross-Lingual Confidence Estimation for Language Models [14.815594636224747]
  We use a lightweight linear probe that predicts answer correctness directly from intermediate representations.<n> Learned layer weights and multiple ablations reveal that confidence features concentrate in middle layers across languages.<n>While zero-shot cross-lingual performance depends on similarity to the source language, the probe provides a strong baseline without any retraining.
  arXiv  Detail & Related papers  (2026-05-29T12:25:24Z)
• Evaluating Metalinguistic Knowledge in Large Language Models across the World's Languages [0.5949779668853555]
  Large language models (LLMs) are routinely evaluated on language use tasks, yet their knowledge of linguistic structure remains poorly understood.<n>Using accuracy and macro F1, together with majority-class and chance baselines, we analyse overall performance and examine variation by linguistic domains and language-related factors.<n>Our results show that metalinguistic knowledge in current LLMs is limited: GPT-4o performs best but only moderate accuracy (0.367), while open-source models lag behind.
  arXiv  Detail & Related papers  (2026-02-02T14:49:56Z)
• Investigating the Multilingual Calibration Effects of Language Model Instruction-Tuning [58.355275813623685]
  This work looks at a critical gap in the calibration of large language models (LLMs) within multilingual settings.<n>Even in low-resource languages, model confidence can increase significantly after instruction-tuning on high-resource language SFT datasets.<n>However, improvements in accuracy are marginal or non-existent, highlighting a critical shortcoming of standard SFT for multilingual languages.
  arXiv  Detail & Related papers  (2026-01-04T04:29:12Z)
• Humans overrely on overconfident language models, across languages [32.71245803698373]
  We study the risks of multilingual linguistic (mis)calibration, overconfidence, and overreliance across five languages.<n>Our work finds that overreliance risks are high across languages.
  arXiv  Detail & Related papers  (2025-07-08T18:01:01Z)
• When Less Language is More: Language-Reasoning Disentanglement Makes LLMs Better Multilingual Reasoners [111.50503126693444]
  We show that language-specific ablation consistently boosts multilingual reasoning performance.<n>Compared to post-training, our training-free ablation achieves comparable or superior results with minimal computational overhead.
  arXiv  Detail & Related papers  (2025-05-21T08:35:05Z)
• PolyMath: Evaluating Mathematical Reasoning in Multilingual Contexts [85.78821098963607]
  PolyMath is a multilingual mathematical reasoning benchmark covering 18 languages and 4 easy-to-hard difficulty levels.<n>Our benchmark ensures difficulty comprehensiveness, language diversity, and high-quality translation.
  arXiv  Detail & Related papers  (2025-04-25T15:39:04Z)
• The Lou Dataset -- Exploring the Impact of Gender-Fair Language in German Text Classification [57.06913662622832]
  Gender-fair language fosters inclusion by addressing all genders or using neutral forms. Gender-fair language substantially impacts predictions by flipping labels, reducing certainty, and altering attention patterns. While we offer initial insights on the effect on German text classification, the findings likely apply to other languages.
  arXiv  Detail & Related papers  (2024-09-26T15:08:17Z)
• Understanding and Mitigating Language Confusion in LLMs [76.96033035093204]
  We evaluate 15 typologically diverse languages with existing and newly-created English and multilingual prompts.<n>We find that Llama Instruct and Mistral models exhibit high degrees of language confusion.<n>We find that language confusion can be partially mitigated via few-shot prompting, multilingual SFT and preference tuning.
  arXiv  Detail & Related papers  (2024-06-28T17:03:51Z)
• Boosting Cross-Lingual Transfer via Self-Learning with Uncertainty Estimation [34.97086123805344]
  Recent multilingual pre-trained language models have achieved remarkable zero-shot performance. We propose a self-learning framework that further utilizes unlabeled data of target languages. We evaluate our framework with uncertainties on two cross-lingual tasks including Named Entity Recognition (NER) and Natural Language Inference (NLI) covering 40 languages in total.
  arXiv  Detail & Related papers  (2021-09-01T05:26:46Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.