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

TL;DR

Disentangling language and reasoning representations in large language models through causal intervention improves multilingual reasoning performance without additional training, inspired by cognitive neuroscience insights.

Contribution

This paper introduces a training-free method to enhance multilingual reasoning in LLMs by ablating language-specific components, demonstrating effectiveness across diverse languages.

Findings

Language-reasoning representations can be effectively disentangled in LLMs.

Ablating language-specific features improves multilingual reasoning performance.

The method maintains linguistic fidelity while boosting reasoning capabilities.

Abstract

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.