Enhancing Multilingual Reasoning via Steerable Model Merging

2026-06-17

AI Summary

This work addresses the limitations of traditional model fusion approaches, which rely on static strategies and struggle to dynamically reconcile conflicts between multilingual and reasoning models during multilingual inference, thereby constraining overall performance. To overcome this, the authors propose ST-Merge, a controllable fusion framework that incorporates a gated cross-attention mechanism to enable input-aware adaptive weight modulation, dynamically balancing the contributions of the two source models. By transcending the constraints of static fusion, ST-Merge achieves significant improvements over multiple strong baselines across four multilingual reasoning benchmarks spanning 21 languages, demonstrating exceptional generalization capability and reasoning performance.

Abstract

Model merging is an effective technique for composing the capabilities of a multilingual model and a reasoning model. It has achieved promising generalization in multilingual reasoning tasks by aligning feature spaces of different models. However, the merged single model often fails to address the conflicts between source models, leading to suboptimal performance. In other words, the one-size-fits-all merging strategy may not align with the characteristics of different inputs which may require prioritizing certain models over others. To this end, we propose a Steerable Model Merging (ST-Merge) framework to modulate the contribution of each source model. To realize this idea, we introduce a gated cross-attention mechanism to weight or filter the two attended source models in an adaptive manner. Extensive experiments demonstrate that ST-Merge consistently outperforms multiple strong baselines on four multilingual reasoning benchmarks across 21 different languages.

Problem

• multilingual reasoning
• model merging
• conflict resolution
• adaptive contribution
• feature alignment

Innovation

• Steerable Model Merging
• gated cross-attention
• multilingual reasoning
• model composition
• adaptive fusion