Decision-Level Ordinal Modeling for Multimodal Essay Scoring with Large Language Models

Automated essay scoring using large language models (LLMs) typically involves autoregressive token generation, where the final score is obtained through decoding and parsing, making the decision-making process implicit. However, this approach is particularly problematic in multimodal essay scoring, where visual information is also considered. A new approach, decision-level ordinal modeling, has been proposed to address this issue by making the decision-making process explicit. This method allows for the prediction of multiple rubric-defined trait scores for each essay, taking into account the ordered discrete rating scale of each trait. The use of LLMs in this context has significant implications for security, as developments in this area can both enhance capabilities and introduce new risks¹. As LLMs continue to evolve, their potential impact on various applications, including those in the DeFi space, will be shaped by their ability to effectively and securely process multimodal data. This development matters to practitioners because it highlights the need to carefully consider the security implications of LLMs in high-stakes applications.