# Language Drift in Multilingual Retrieval-Augmented Generation: Characterization and Decoding-Time Mitigation

arXiv: https://arxiv.org/abs/2511.09984 · HTML: https://mlsanity.com/paper/2511.09984/

## Key facts

- Cited by 8 papers in the map (data through 14 Oct 2026).
- Placed in Generation and context integration, one of 15 directions in Retrieval-augmented generation.
- Builds on 4 of its 22 references that are in the map.
- First posted on arXiv 13 Nov 2025.

## Placed in

- Retrieval-augmented generation / [Generation and context integration](https://mlsanity.com/rag/generation-and-context-integration/)

## Builds on (most cited first, 4 of 4)

- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [Qwen3 Technical Report](https://mlsanity.com/paper/2505.09388/) (2025)
- [Qwen2.5 Technical Report](https://mlsanity.com/paper/2412.15115/) (2024)
- [Generate-then-Ground in Retrieval-Augmented Generation for Multi-hop Question Answering](https://mlsanity.com/paper/2406.14891/) (2024)

## Built on by (most cited first, 8 of 8)

- [Retrieval as Generation: A Unified Framework with Self-Triggered Information Planning](https://mlsanity.com/paper/2604.11407/) (2026)
- [Enhancing Multilingual RAG Systems with Debiased Language Preference-Guided Query Fusion](https://mlsanity.com/paper/2601.02956/) (2026)
- [Learning from AVA: Early Lessons from a Curated and Trustworthy Generative AI for Policy and Development Research](https://mlsanity.com/paper/2604.17843/) (2026)
- [CroSearch-R1: Better Leveraging Cross-lingual Knowledge for Retrieval-Augmented Generation](https://mlsanity.com/paper/2604.25182/) (2026)
- [CORAL: Adaptive Retrieval Loop for Culturally-Aligned Multilingual RAG](https://mlsanity.com/paper/2604.25676/) (2026)
- [Distill Where the Student Goes: Teacher-Regularized RL for English-Evidence Cross-Lingual RAG](https://mlsanity.com/paper/2607.02966/) (2026)
- [When Simpler Is Better: Evaluating Translation Pipelines for Medieval Latin Manuscripts](https://mlsanity.com/paper/2607.03836/) (2026)
- [LLMs Get Smarter from Targeted Synthetic Multilingual Data](https://mlsanity.com/paper/2608.15964/) (2026)

## Abstract (arXiv)

Multilingual Retrieval-Augmented Generation (RAG) enables large language models (LLMs) to perform knowledge-intensive tasks in multilingual settings by leveraging retrieved documents as external evidence. However, when the retrieved evidence differs in language from the user query and in-context exemplars, the model often exhibits language drift by generating responses in an unintended language. This phenomenon is especially pronounced during reasoning-intensive decoding, such as Chain-of-Thought (CoT) generation, where intermediate steps introduce further language instability. In this paper, we systematically study output language drift in multilingual RAG across multiple datasets, languages, and LLM backbones. Our controlled experiments reveal that the drift results not from comprehension failure but from decoder-level collapse, where dominant token distributions and high-frequency English patterns dominate the intended generation language. We further observe that English serves as a semantic attractor under cross-lingual conditions, emerging as both the strongest interference source and the most frequent fallback language. To mitigate this, we propose Soft Constrained Decoding (SCD), a lightweight, training-free decoding strategy that gently steers generation toward the target language by penalizing non-target-language tokens. SCD is model-agnostic and can be applied to any generation algorithm without modifying the architecture or requiring additional data. Experiments across three multilingual datasets and multiple typologically diverse languages show that SCD consistently improves language alignment and task performance, providing an effective and generalizable solution in multilingual RAG.