LLM Judges as Raters: A Pre-Registered Audit of Severity, Halo, Reliability, and Version Instability in LLM Essay Scoring on Public Corpora

Despite growing interest in using Large Language Models (LLMs) for educational assessment, it remains unclear how closely they align with human scoring. We present a systematic evaluation of instruction-tuned LLMs across three open essay-scoring datasets (ASAP 2.0, ELLIPSE, and DREsS) that cover both holistic and analytic scoring. We analyze agreement with human consensus scores, directional bias, and the stability of bias estimates. Our results show that strong open-weight models achieve moderate to high agreement with humans on holistic scoring (Quadratic Weighted Kappa about 0.6), but this does not transfer uniformly to analytic scoring. In particular, we observe large and stable negative directional bias on Lower-Order Concern (LOC) traits, such as Grammar and Conventions, meaning that models often score these traits more harshly than human raters. We also find that concise keyword-based prompts generally outperform longer rubric-style prompts in multi-trait analytic scoring. To quantify the amount of data needed to detect these systematic deviations, we compute the minimum sample size at which a 95% bootstrap confidence interval for the mean bias excludes zero. This analysis shows that LOC bias is often detectable with very small validation sets, whereas Higher-Order Concern (HOC) traits typically require much larger samples. These findings support a bias-correction-first deployment strategy: instead of relying on raw zero-shot scores, systematic score offsets can be estimated and corrected using small human-labeled bias-estimation sets, without requiring large-scale fine-tuning.

Large language models have recently been proposed as tools for automated essay scoring, but their agreement with human grading remains unclear. In this work, we evaluate how LLM-generated scores compare with human grades and analyze the grading behavior of several models from the GPT and Llama families in an out-of-the-box setting, without task-specific training. Our results show that agreement between LLM and human scores remains relatively weak and varies with essay characteristics. In particular, compared to human raters, LLMs tend to assign higher scores to short or underdeveloped essays, while assigning lower scores to longer essays that contain minor grammatical or spelling errors. We also find that the scores generated by LLMs are generally consistent with the feedback they generate: essays receiving more praise tend to receive higher scores, while essays receiving more criticism tend to receive lower scores. These results suggest that LLM-generated scores and feedback follow coherent patterns but rely on signals that differ from those used by human raters, resulting in limited alignment with human grading practices. Nevertheless, our work shows that LLMs produce feedback that is consistent with their grading and that they can be reliably used in supporting essay scoring.

Researchers increasingly use Large Language Models as judges (LLM-as-a-judge) to evaluate model outputs. Yet there are no standards for how to design these judges. Typically, researchers choose the prompt, rating scale, and model intuitively. If these choices change the judge's verdicts, two studies can reach different conclusions about the same facts. To address this risk and to provide an empirical basis for judge designs, we evaluate 10 reasoning models across multiple designs on two tasks: a scalar rating of sentence sentiment and toxicity (over 500 items per category), as well as a binary accuracy classification of question-answer pairs (n=600). For the rating tasks, despite judges showing significant disagreements with the human ground truth, the practical size of differences is small enough to consider most judges reliable (mean absolute deviation of 0.11 points on a 1 - 7 scale); toxicity judges even outperform standard classifiers. Judges are also highly accurate on average (96.5%) for the accuracy classification task. However, design choices can produce shifts: changing the rating scale alone can shift measured bias by up to 0.93 points (rating task), and while accuracy levels are rarely impacted, design choices consistently impact judge leniency (classification task; leniency drop of 28.9 percentage points when using detailed prompts, and up to 56.1 percentage points when switching models). Counterintuitively, lower reasoning effort affects neither accuracy nor leniency. Across both tasks, model identity is the dominant source of variance. These findings suggest that while LLM judges are broadly trustworthy in aggregate, design choices can be meaningful sources of variance. Given the growing reliance on automated evaluation in LLM research, we intend this study as a methodological reference for designing more robust and replicable LLM-as-a-judge pipelines.