LANG: Reinforcement Learning for Multilingual Reasoning with Language-Adaptive Hint Guidance

TL;DR

The paper introduces LANG, a reinforcement learning framework that improves multilingual reasoning in large language models by using language-conditioned hints and adaptive strategies to balance reasoning quality and language consistency.

Contribution

LANG is a novel framework that employs language-adaptive hint guidance and decay schedules to enhance multilingual reasoning without losing language fidelity.

Findings

LANG significantly improves reasoning performance on multilingual benchmarks.

The framework maintains language consistency while boosting reasoning accuracy.

LANG generalizes beyond mathematics to improve language alignment across model layers.

Abstract

Reinforcement learning has proven effective for enhancing multi-step reasoning in large language models (LLMs), yet its benefits have not fully translated to multilingual contexts. Existing methods struggle with a fundamental trade-off: prioritizing input-language consistency severely hampers reasoning quality, while prioritizing reasoning often leads to unintended language drift toward English. We address this challenge with LANG, a novel framework that leverages language-conditioned hints to guide exploration in non-English reasoning tasks. Our method incorporates two key mechanisms to prevent dependency on these hints: a progressive decay schedule that gradually withdraws scaffolding, and a language-adaptive switch that tailors learning horizons to specific language difficulties. Empirical results on challenging multilingual mathematical benchmarks reveal that LANG substantially enhances reasoning performance without compromising language consistency. Moreover, we show that our framework generalizes beyond mathematics, fostering more consistent language alignment across model layers