CONGRAD: Conflicting Gradient Filtering for Multilingual Preference Alignment

AbstractNaive joint training of large language models (LLMs) for multilingual preference alignment can suffer from negative interference. This is a known issue in multilingual training, where conflicting objectives degrade overall performance. However, the impact of this phenomenon in the context of multilingual preference alignment remains largely underexplored. To address this issue, we propose ConGrad, an effective and scalable filtering method that mitigates this interference by identifying and selecting preference samples that exhibit high cross-lingual affinity. Based on principles of multi-objective optimization, our approach computes an aggregated, cross-lingually beneficial gradient direction and uses this to filter for samples whose individual gradients align with this consensus direction. To ensure scalability for LLMs, we incorporate a sublinear gradient compression strategy that reduces memory overhead during gradient accumulation. We integrate ConGrad into a self-rewarding framework and evaluate on LLaMA3-8B and Gemma2-2B across 10 languages. Results show that ConGrad consistently outperforms strong baselines in both seen and unseen languages, with minimal alignment tax.