Steering Merged LLMs for Multilingual Reasoning with Coefficient Optimization

Zhuoran Li, Rui Xu, Xiaojie Wang, Junnan Liu, Zhijun Chen, Qianren Mao, Jian Yang, Hongcheng Guo, Likang Xiao, MING LI

Keywords: multilingual reasoning, LLM, steering, model merging

TL;DR: Steering Merged LLMs for Multilingual Reasoning with Coefficient Optimization

Abstract: Model merging has recently proven effective in enhancing the cross-lingual capabilities of reasoning models by integrating them with multilingual language models. However, existing methods typically apply a uniform merging strategy across languages, leading to a trade-off: while low-resource languages may benefit, high-resource languages such as English often suffer performance degradation. We attribute this limitation to insufficient coordination between the multilingual and reasoning models, where suboptimal representation merging impairs generalization. To mitigate this, we introduce **SteMerger**, a preference-driven framework that dynamically **steers** the model **merger** by optimizing the merging coefficients. Experiments on multilingual reasoning benchmarks show that SteMerger consistently improves performance across a wide range of languages, outperforming several strong baselines.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 10691