Cross-Preference Learning for Sentence-Level and Context-Aware Machine Translation

To address the problem of uneven signal benefits in context-aware machine translation, researchers proposed the Cross-Preference Learning (CPL) framework. This framework integrates intra-sentence and inter-sentence preferences into the optimization objectives, enabling explicit utilization of informational contexts while maintaining robustness against non-informational contexts. Experiments were conducted on multiple public tasks, and the models tested included Qwen3-4B, Qwen3-8B, and Llama-3-8B-Instruct. The results showed that without any structural modifications, the CPL framework achieved continuous improvement in translation quality and robustness under both input conditions.