A comprehensive analysis of evaluating robustness and generalization ability of models in AES

Abstract Large language models are capable of impressive performance in standard benchmark testing, but are often fragile when deployed in real-world scenarios and cannot be deployed in many scenarios directly, such as navigation and autonomous driving. When some adversarial samples require continuous learning of new data distributions, the robustness and generalization abilities of the model require detailed evaluations and further mining. Therefore, it is very necessary to use tools to evaluate the robustness and generalization ability of the model before deploying models in real-world scenarios. The robustness and generalization capabilities of models were extensively assessed before their implementation in Automatic essay Scoring. This was achieved through a comprehensive series of experiments aimed at validating the efficacy of various models. Our findings underscored the significance of different models and their impact, providing valuable insights for practitioners seeking to forecast the performance of deployed models across diverse distributions.