R0027/2026-03-26/Q001/S02/R04¶
Multilingual LLM performance biases across 9 languages in educational tasks
Summary¶
| Field | Value |
|---|---|
| Title | Multilingual Performance Biases of Large Language Models in Education |
| URL | https://arxiv.org/html/2504.17720v2 |
| Date accessed | 2026-03-26 |
| Publication date | 2025-04 |
| Author(s) | Vansh Gupta, Sankalan Pal Chowdhury, Vilem Zouhar, Donya Rooein, Mrinmaya Sachan |
| Publication | arXiv preprint (ETH Zurich / Bocconi) |
Selection Decision¶
Included in evidence base: Yes
Rationale: Covers 9 languages including Hindi, Mandarin, and Arabic with per-language accuracy data. Includes the critical finding that English prompts outperform translated prompts (72.7% vs 67.2%).