English

Aligning Multilingual Reasoning with Verifiable Semantics from a High-Resource Expert Model

Computation and Language 2025-10-01 v1 Artificial Intelligence

Abstract

While reinforcement learning has advanced the reasoning abilities of Large Language Models (LLMs), these gains are largely confined to English, creating a significant performance disparity across languages. To address this, we introduce Pivot-Based Reinforcement Learning with Semantically Verifiable Rewards (PB-RLSVR), a novel framework that enhances multilingual reasoning by circumventing the need for human-annotated data in target languages. Our approach employs a high-performing English LLM as a "pivot" model to generate reference responses for reasoning tasks. A multilingual model is then rewarded based on the semantic equivalence of its responses to the English reference, effectively transferring the pivot model's reasoning capabilities across languages. We investigate several cross-lingual semantic reward functions, including those based on embeddings and machine translation. Extensive experiments on a suite of multilingual reasoning benchmarks show that our method significantly narrows the performance gap between English and other languages, substantially outperforming traditional PPO baselines. Specifically, our PB-RLSVR framework improves the average multilingual performance of Llama-3.1-8B-Instruct and Qwen3-32B by 16.41% and 10.17%, respectively, demonstrating a powerful and data-efficient approach to building truly multilingual reasoning agents.

Keywords

Cite

@article{arxiv.2509.25543,
  title  = {Aligning Multilingual Reasoning with Verifiable Semantics from a High-Resource Expert Model},
  author = {Fahim Faisal and Kaiqiang Song and Song Wang and Simin Ma and Shujian Liu and Haoyun Deng and Sathish Reddy Indurthi},
  journal= {arXiv preprint arXiv:2509.25543},
  year   = {2025}
}
R2 v1 2026-07-01T06:06:22.191Z