Research map / Large language models / Retrieval and knowledge
Language Drift in Multilingual Retrieval-Augmented Generation: Characterization and Decoding-Time Mitigation
Bo Li, Zhenghua Xu, Rui Xie
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 Soft Constrained Decoding (boost target tokens, penalize distractor tokens) | LLaMA3-8B-Instruct · HotpotQA (ZH target, EN context) | language consistency: 68.4% -> 90.6% | Improved |
| add SCD | LLaMA3-8B-Instruct · HotpotQA (ZH target, EN context) | ROUGE: 0.182 -> 0.306 | Improved |
| add SCD | LLaMA3-8B-Instruct · MuSiQue (ZH target, EN context) | language consistency: 63.9% -> 91.8% | Improved |
| add SCD instead of translating outputs back after generation (ablation) | LLaMA3-8B-Instruct · HotpotQA (ZH target, EN context) | ROUGE: 0.263 -> 0.306 | Improved |
| use soft penalty (SCD) instead of hard vocabulary restriction (ablation) | LLaMA3-8B-Instruct · HotpotQA (ZH target, EN context) | ROUGE: 0.155 -> 0.306 | Improved |
| hard vocabulary restriction instead of prompted instruction (ablation) | LLaMA3-8B-Instruct · HotpotQA (AR/RU targets) | ROUGE: 0.294 -> 0.295 (AR); 0.333 -> 0.343 (RU) | No clear effect |
| hard vocabulary restriction instead of prompted instruction (ablation) | LLaMA3-8B-Instruct · HotpotQA (ZH target, EN context) | ROUGE: 0.182 -> 0.155 | Hurt |
| switch retrieved context from ZH to EN (cross-lingual interference) | LLaMA3-8B-Instruct with ICL · HotpotQA (ZH target) | language consistency: 92.0% -> 68.4% | Hurt |
| add in-context exemplars | LLaMA3-8B-Instruct · HotpotQA (RU target) | ROUGE and language consistency: ROUGE 0.193 -> 0.373; LC 0.991 -> 0.895 | Mixed |
| translate drifted outputs back to the target language before scoring | LLaMA3-8B-Instruct · HotpotQA (ZH target, EN context) | ROUGE: 0.182 -> 0.263 | Improved |
Closest papers by meaning
- Benchmarking Large Language Models in Retrieval-Augmented Generation (AAAI 2024 · similarity 0.91)
- Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG (ICLR 2025 · similarity 0.90)
- RAGGED: Towards Informed Design of Scalable and Stable RAG Systems (ICML 2025 · similarity 0.90)
- A Theory for Token-Level Harmonization in Retrieval-Augmented Generation (ICLR 2025 · similarity 0.90)
- Inference Scaling for Long-Context Retrieval Augmented Generation (ICLR 2025 · similarity 0.89)
- RAPID: Long-Context Inference with Retrieval-Augmented Speculative Decoding (ICML 2025 · similarity 0.89)
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.
