Skewed Score: A statistical framework to assess autograders

2025-07-04

AI Summary

While large language model (LLM)-based automatic scorers offer scalability, their reliability is low and bias sources are multifaceted—spanning response type, domain, and scoring methodology—necessitating joint assessment of their internal reliability and bias. Method: We propose a Bayesian generalized linear model (GLM) framework that jointly models rater characteristics (e.g., LLM architecture) and response attributes (e.g., length, generating model) to simultaneously quantify scoring system performance, systematic biases, and uncertainty. Contribution/Results: Unlike conventional reliability metrics, our approach enhances interpretability, enables fine-grained bias attribution, and yields posterior probability estimates for bias differences. Empirical evaluation across multi-dimensional scoring tasks demonstrates its effectiveness in bias detection and reliability calibration.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Relational Probabilistic Models
• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

The evaluation of large language model (LLM) outputs is increasingly performed by other LLMs, a setup commonly known as "LLM-as-a-judge", or autograders. While autograders offer a scalable alternative to human evaluation, they have shown mixed reliability and may exhibit systematic biases, depending on response type, scoring methodology, domain specificity, and other factors. In this paper we propose a statistical framework based on Bayesian generalised linear models (GLMs) that enables researchers to simultaneously assess their autograders while also addressing their primary research questions (e.g., LLM evaluation). Our approach models evaluation outcomes (e.g., scores or pairwise preferences) as a function of properties of the grader (e.g., human vs. autograder) and the evaluated item (e.g., response length or the LLM that generated it), allowing for explicit quantification of scoring differences and potential biases within a unified framework. In addition, our method can be used to augment traditional reliability metrics such as inter-rater agreement, by providing uncertainty estimates and clarifying the source of disagreement. Overall, this approach contributes to more robust and interpretable use of autograders in LLM evaluation, enabling both performance analysis and bias detection.

Problem

• Assess reliability and biases in LLM-based autograders
• Quantify scoring differences between human and autograders
• Enhance interpretability of autograder performance and biases

Innovation

• Bayesian GLMs for autograder assessment
• Quantifies scoring differences and biases
• Augments reliability metrics with uncertainty