Do Large Language Models Judge Error Severity Like Humans?

2025-06-05

AI Summary

This study investigates the alignment between large language models (LLMs) and human judgment in assessing the severity of semantic errors in image captions—specifically age, gender, clothing type, and color errors. Building upon van Miltenburg’s experimental framework, we design a controlled error-injection paradigm, conduct multimodal versus unimodal comparative evaluations, and analyze underlying neurocognitive mechanisms. Our key contributions are threefold: (1) We systematically identify social norm bias in LLMs—e.g., over-penalizing gender errors—and perceptual misalignment—e.g., overestimating color-error severity; (2) We demonstrate that visual context substantially enhances human sensitivity to color and type errors, whereas most LLMs exhibit significant deviations from human severity rankings; (3) Among tested models, Doubao shows partial alignment but weak discriminative power, while the unimodal DeepSeek-V3 achieves the highest human alignment across all settings—surpassing multimodal counterparts.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Language and Vision
• Natural Language Processing: (Large) Language Models

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Large Language Models (LLMs) are increasingly used as automated evaluators in natural language generation, yet it remains unclear whether they can accurately replicate human judgments of error severity. In this study, we systematically compare human and LLM assessments of image descriptions containing controlled semantic errors. We extend the experimental framework of van Miltenburg et al. (2020) to both unimodal (text-only) and multimodal (text + image) settings, evaluating four error types: age, gender, clothing type, and clothing colour. Our findings reveal that humans assign varying levels of severity to different error types, with visual context significantly amplifying perceived severity for colour and type errors. Notably, most LLMs assign low scores to gender errors but disproportionately high scores to colour errors, unlike humans, who judge both as highly severe but for different reasons. This suggests that these models may have internalised social norms influencing gender judgments but lack the perceptual grounding to emulate human sensitivity to colour, which is shaped by distinct neural mechanisms. Only one of the evaluated LLMs, Doubao, replicates the human-like ranking of error severity, but it fails to distinguish between error types as clearly as humans. Surprisingly, DeepSeek-V3, a unimodal LLM, achieves the highest alignment with human judgments across both unimodal and multimodal conditions, outperforming even state-of-the-art multimodal models.

Problem

• Assessing if LLMs replicate human error severity judgments
• Comparing human and LLM evaluations of semantic errors
• Exploring LLM biases in gender versus color error severity

Innovation

• Systematically compare human and LLM assessments
• Extend framework to unimodal and multimodal settings
• Evaluate four error types with controlled errors