A rigorous psychometric approach is crucial for the accurate measurement of mind-reading abilities. Traditional scoring methods for such tests, which involve lengthy free-text responses, require considerable time and human effort. This study investigates the use of large language models (LLMs) to automate the scoring of psychometric tests. Data were collected from participants aged 13 to 30 years and scored by trained human coders to establish a benchmark. We evaluated multiple LLMs against human assessments, exploring various prompting strate- gies to optimize performance and fine-tuning the models using a subset of the collected data to enhance accuracy. Our results demonstrate that LLMs can assess advanced mind-reading abilities with over 90% accuracy on average. Notably, in most test items, the LLMs achieved higher Kappa agreement with the lead coder than two trained human coders, highlighting their potential to reliably score open-response psychometric tests.

In this paper we study the patterns of label disagreement in data used for instruction tuning Large Language models (LLMs). Specifically, we focus on data used for Reinforcement Learning from Human Feedback (RLHF). Our objective is to determine what is the primary source of disagreement: the individual data points, the choice of annotators, or the task formulation. We annotate the same dataset multiple times under different conditions and compare the overall agreement and the patterns of disagreement. For task formulation, we compare “single” format where annotators rate LLM responses individually with “preference” format where annotators select one of two possible responses. For annotators, we compare data from human labelers with automatic data labeling using LLMs. Our results indicate that: (1) there are very few “universally ambiguous” instances. The label disagreement depends largely on the task formulation and the choice of annotators; (2) the overall agreement remains consistent across experiments. We find no evidence that “preference” data is of higher quality than “single” data; and (3) the change of task formulation and annotators impacts the resulting instance-level labels. The labels obtained in different experiments are correlated, but not identical.