<p>Automated short answer scoring (SAS) is the task of automatically scoring a given input to a prompt based on rubrics and reference answers. SAS is promising for real-world applications. However, because rubrics and reference answers differ among prompts, there is a need to acquire new data and train a model for each new prompt. This makes SAS expensive, especially in schools and online courses where resources are limited and only a few prompts are used. In this study, we propose a two-phase approach to address this issue. The proposed approach involves training a model on existing rubrics and answers with gold score signals and then finetuning the model on a new prompt. In particular, given that scoring rubrics and reference answers differ for different prompts, we employed key phrases, which are representative expressions that the answer should contain to gain a score, and trained an SAS model to learn the relationship between the key phrases and answers using already annotated prompts (i.e., cross-prompts). We evaluated the proposed approach using bidirectional encoder representations from transformers (BERT) and open-source large language models (LLMs). In addition, we incorporated the proposed approach with zero-shot conditions and in-context learning of LLMs. The results show that the proposed two-phase approach significantly improves scoring accuracy, especially when the training data is limited. Finally, an extensive analysis revealed that it is crucial to design a model that can learn a task’s general properties.</p>