Research map / Large language models / Retrieval and knowledge
Linguistic Nepotism: Trading-off Quality for Language Preference in Multilingual RAG
Dayeon Ki, Marine Carpuat, Paul McNamee, Daniel Khashabi, Eugene Yang, Dawn Lawrie, Kevin Duh
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 |
|---|---|---|---|
| translate the cited document from English to Swahili (ablation) | six open-weight LLMs · ELI5 citation prediction | citation accuracy (%): AYA23 8B: 60.0 -> 22.4; Qwen3 8B: 62.6 -> 30.4 | Hurt |
| translate the cited document from English to Bengali (ablation) | six open-weight LLMs · ELI5 citation prediction | citation accuracy (%): AYA23 8B: 60.0 -> 27.2 | Hurt |
| translate the cited document from English to French or Spanish (ablation) | six open-weight LLMs · ELI5 citation prediction | citation accuracy (%): average -8.82% (French), -8.08% (Spanish) | Hurt |
| place the cited document in the middle of the context (ablation) | six open-weight LLMs · ELI5 citation prediction | English preference gap: largest accuracy gap vs English in the middle position | Hurt |
| increase model size (ablation) | Llama/Qwen3/Gemma-3 models · ELI5 English citation accuracy | citation accuracy (%): Qwen3 8B 62.6 -> Qwen3 14B 83.0 | Improved |
| use a multilingual-trained model (Aya23 8B) (ablation) | Aya23 8B · ELI5 citation prediction | English preference | No clear effect |
| query in target language, cited document in query language and other documents in English (ablation) | six open-weight LLMs · ELI5, translated queries | citation accuracy: highest accuracy in 28 of 48 model-language pairs | Improved |
| make all documents match the query language instead of only the cited one (ablation) | six open-weight LLMs · ELI5, translated queries | citation accuracy: language contrast variant generally beats all-query-language variant | No clear effect |
| put relevant document in target language with irrelevant English distractor (ablation) | six open-weight LLMs · MIRACL English dev (231 queries) | citation accuracy | Hurt |
| put irrelevant distractor in target language, relevant document in English (ablation) | six open-weight LLMs · MIRACL English dev (231 queries) | citation accuracy | Improved |
Closest papers by meaning
- Chain-of-Thought Improves Text Generation with Citations in Large Language Models (AAAI 2024 · similarity 0.85)
- SePer: Measure Retrieval Utility Through The Lens Of Semantic Perplexity Reduction (ICLR 2025 · similarity 0.84)
- HLM-Cite: Hybrid Language Model Workflow for Text-based Scientific Citation Prediction (NeurIPS 2024 · similarity 0.83)
- A Theory for Token-Level Harmonization in Retrieval-Augmented Generation (ICLR 2025 · similarity 0.83)
- Benchmarking Large Language Models in Retrieval-Augmented Generation (AAAI 2024 · similarity 0.83)
- KnowPO: Knowledge-Aware Preference Optimization for Controllable Knowledge Selection in Retrieval-Augmented Language Models (AAAI 2025 · similarity 0.83)
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.
