The Impact of Language Mixing on Bilingual LLM Reasoning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Li, Yihao, Xin, Jiayi, Miao, Miranda Muqing, Long, Qi, Ungar, Lyle
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918151235567616
author Li, Yihao
Xin, Jiayi
Miao, Miranda Muqing
Long, Qi
Ungar, Lyle
author_facet Li, Yihao
Xin, Jiayi
Miao, Miranda Muqing
Long, Qi
Ungar, Lyle
contents Proficient multilingual speakers often intentionally switch languages in the middle of a conversation. Similarly, recent reasoning-focused bilingual large language models (LLMs) with strong capabilities in both languages exhibit language mixing-alternating languages within their chain of thought. Discouraging this behavior in DeepSeek-R1 was found to degrade accuracy, suggesting that language mixing may benefit reasoning. In this work, we study language switching in Chinese-English bilingual reasoning models. We identify reinforcement learning with verifiable rewards (RLVR) as the critical training stage that leads to language mixing. We show that language mixing can enhance reasoning: enforcing monolingual decoding reduces accuracy by 5.6 percentage points on MATH500. Additionally, a lightweight probe can be trained to predict whether a potential language switch would benefit or harm reasoning, and when used to guide decoding, increases accuracy by 2.92 percentage points. Our findings suggest that language mixing is not merely a byproduct of multilingual training, but is a strategic reasoning behavior.
format Preprint
id arxiv_https___arxiv_org_abs_2507_15849
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle The Impact of Language Mixing on Bilingual LLM Reasoning
Li, Yihao
Xin, Jiayi
Miao, Miranda Muqing
Long, Qi
Ungar, Lyle
Computation and Language
Artificial Intelligence
Machine Learning
Proficient multilingual speakers often intentionally switch languages in the middle of a conversation. Similarly, recent reasoning-focused bilingual large language models (LLMs) with strong capabilities in both languages exhibit language mixing-alternating languages within their chain of thought. Discouraging this behavior in DeepSeek-R1 was found to degrade accuracy, suggesting that language mixing may benefit reasoning. In this work, we study language switching in Chinese-English bilingual reasoning models. We identify reinforcement learning with verifiable rewards (RLVR) as the critical training stage that leads to language mixing. We show that language mixing can enhance reasoning: enforcing monolingual decoding reduces accuracy by 5.6 percentage points on MATH500. Additionally, a lightweight probe can be trained to predict whether a potential language switch would benefit or harm reasoning, and when used to guide decoding, increases accuracy by 2.92 percentage points. Our findings suggest that language mixing is not merely a byproduct of multilingual training, but is a strategic reasoning behavior.
title The Impact of Language Mixing on Bilingual LLM Reasoning
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2507.15849