Zero-shot relation triplet extraction (ZeroRTE) aims to extract structured triplets from text without training on target relations. While large language models show promise for zero-shot tasks, existing approaches often fail to equip models with sufficient discriminative capability, leading to imprecise relation predictions and limited generalization in zero-shot scenarios. We propose Relation Discrimination Learning (RDL), a simple yet effective framework that reframes relation extraction as a discrimination task. Instead of direct extraction, RDL trains models to select the correct relation from a candidate set, thereby enhancing the model’s ability to make fine-grained comparisons. This approach directly addresses the core need for improved relation discrimination. Extensive experiments show that RDL achieves significant and consistent improvements over state-of-the-art baselines across different model scales and datasets. It maintains excellent performance in low-resource scenarios while requiring no architectural modifications or multi-stage pipeline design. Our work establishes relation discrimination as a central principle for zero-shot relation triplet extraction, offering a new perspective to the community.