When Less Language is More: Language-Reasoning Disentanglement Makes LLMs Better Multilingual Reasoners

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Zhao, Weixiang, Guo, Jiahe, Deng, Yang, Wu, Tongtong, Zhang, Wenxuan, Hu, Yulin, Sui, Xingyu, Zhao, Yanyan, Che, Wanxiang, Qin, Bing, Chua, Tat-Seng, Liu, Ting
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911311961522176
author Zhao, Weixiang
Guo, Jiahe
Deng, Yang
Wu, Tongtong
Zhang, Wenxuan
Hu, Yulin
Sui, Xingyu
Zhao, Yanyan
Che, Wanxiang
Qin, Bing
Chua, Tat-Seng
Liu, Ting
author_facet Zhao, Weixiang
Guo, Jiahe
Deng, Yang
Wu, Tongtong
Zhang, Wenxuan
Hu, Yulin
Sui, Xingyu
Zhao, Yanyan
Che, Wanxiang
Qin, Bing
Chua, Tat-Seng
Liu, Ting
contents Multilingual reasoning remains a significant challenge for large language models (LLMs), with performance disproportionately favoring high-resource languages. Drawing inspiration from cognitive neuroscience, which suggests that human reasoning functions largely independently of language processing, we hypothesize that LLMs similarly encode reasoning and language as separable components that can be disentangled to enhance multilingual reasoning. To evaluate this, we perform a causal intervention by ablating language-specific representations at inference time. Experiments on 10 open-weight LLMs spanning 11 typologically diverse languages show that this language-specific ablation consistently boosts multilingual reasoning performance. Layer-wise analyses further confirm that language and reasoning representations can be effectively disentangled throughout the model, yielding improved multilingual reasoning capabilities, while preserving top-layer language features remains essential for maintaining linguistic fidelity. Compared to post-training methods such as supervised fine-tuning or reinforcement learning, our training-free language-reasoning disentanglement achieves comparable or superior results with minimal computational overhead. These findings shed light on the internal mechanisms underlying multilingual reasoning in LLMs and suggest a lightweight and interpretable strategy for improving cross-lingual generalization.
format Preprint
id arxiv_https___arxiv_org_abs_2505_15257
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle When Less Language is More: Language-Reasoning Disentanglement Makes LLMs Better Multilingual Reasoners
Zhao, Weixiang
Guo, Jiahe
Deng, Yang
Wu, Tongtong
Zhang, Wenxuan
Hu, Yulin
Sui, Xingyu
Zhao, Yanyan
Che, Wanxiang
Qin, Bing
Chua, Tat-Seng
Liu, Ting
Computation and Language
Multilingual reasoning remains a significant challenge for large language models (LLMs), with performance disproportionately favoring high-resource languages. Drawing inspiration from cognitive neuroscience, which suggests that human reasoning functions largely independently of language processing, we hypothesize that LLMs similarly encode reasoning and language as separable components that can be disentangled to enhance multilingual reasoning. To evaluate this, we perform a causal intervention by ablating language-specific representations at inference time. Experiments on 10 open-weight LLMs spanning 11 typologically diverse languages show that this language-specific ablation consistently boosts multilingual reasoning performance. Layer-wise analyses further confirm that language and reasoning representations can be effectively disentangled throughout the model, yielding improved multilingual reasoning capabilities, while preserving top-layer language features remains essential for maintaining linguistic fidelity. Compared to post-training methods such as supervised fine-tuning or reinforcement learning, our training-free language-reasoning disentanglement achieves comparable or superior results with minimal computational overhead. These findings shed light on the internal mechanisms underlying multilingual reasoning in LLMs and suggest a lightweight and interpretable strategy for improving cross-lingual generalization.
title When Less Language is More: Language-Reasoning Disentanglement Makes LLMs Better Multilingual Reasoners
topic Computation and Language
url https://arxiv.org/abs/2505.15257