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

2026-05-21 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

LANG is a novel framework addressing the limitations of reinforcement learning (RL) in enhancing multi-step reasoning for large language models (LLMs) in multilingual contexts. Existing RL methods face a fundamental trade-off between maintaining input-language consistency and achieving high reasoning quality, often resulting in unintended language drift toward English. LANG tackles this by using language-conditioned hints to guide exploration in non-English reasoning tasks. It incorporates a progressive decay schedule to gradually withdraw these hints and a language-adaptive switch that tailors learning horizons to specific language difficulties. Empirical results on challenging multilingual mathematical benchmarks demonstrate that LANG substantially improves reasoning performance while preserving language consistency. The framework also generalizes beyond mathematics, promoting more consistent language alignment across model layers.

Key takeaway

For Machine Learning Engineers developing multilingual LLMs, if you are struggling with the trade-off between reasoning quality and language consistency, LANG offers a robust solution. Its language-adaptive hint guidance, combined with progressive decay and tailored learning horizons, significantly enhances non-English reasoning performance without causing language drift. You should explore integrating similar language-conditioned scaffolding techniques to improve your models' performance on diverse multilingual tasks, from mathematics to general reasoning.

Key insights

LANG enhances multilingual LLM reasoning by using language-adaptive hints with a progressive decay schedule, preventing language drift.

Principles

• Multilingual RL requires balancing consistency and reasoning.
• Language-adaptive scaffolding improves non-English reasoning.
• Gradual hint decay prevents over-reliance.

Method

LANG employs language-conditioned hints for non-English reasoning exploration. It uses a progressive decay schedule to withdraw hints and a language-adaptive switch to tailor learning horizons based on language difficulty.

In practice

• Improve LLM reasoning on multilingual math benchmarks.
• Enhance language consistency across model layers.
• Apply RL to diverse non-English reasoning tasks.

Topics

• Reinforcement Learning
• Large Language Models
• Multilingual NLP
• Reasoning
• Language Consistency
• Hint Guidance

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Enhancing Multilingual Reasoning via Steerable Model Merging — Dynamically modulating source model contributions resolves conflicts in multilingual LLM merging for improved reasoning.
• Translate-R1: Cost-Aware Translation Tool Use via Reinforcement Learning — A learned policy allows LLMs to introspect their comprehension and invoke translation tools only when necessary, balancing performance and cost.
• Learning When to Translate for Multilingual Reasoning — RLMs can learn to selectively translate non-English inputs only when direct understanding is unreliable, improving multilingual reasoning.
• Language Effects in Text-to-SQL Across English and Portuguese — Multilingual Text-to-SQL accuracy drops significantly from English to Portuguese, even with advanced LLMs.
• TLPO: Token-Level Policy Optimization for Mitigating Language Confusion in Large Language Models — TLPO uses token-level policy optimization to mitigate language confusion in LLMs without degrading general capabilities.
• Improving Cross-Lingual Factual Recall via Consistency-Driven Reinforcement Learning — GRPO significantly improves cross-lingual factual recall and consistency in LLMs by fostering shared representations, outperforming SFT.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.