Rethinking Ground Truth: A Case Study on Human Label Variation in MLLM Benchmarking

TL;DR

This paper investigates human label variation in MLLM benchmarking, proposing an evaluation protocol that considers both agreement and disagreement among annotators to provide more realistic model assessments.

Contribution

It introduces a new benchmarking protocol for MLLMs that explicitly accounts for human label variation, highlighting the limitations of consensus-based evaluations.

Findings

Larger models perform better on high-agreement tasks.

Medium models often underperform when human disagreement is high.

Benchmarking solely on consensus labels can overestimate model capabilities.

Abstract

Human Label Variation (HLV), i.e. systematic differences among annotators' judgments, remains underexplored in benchmarks despite rapid progress in large language model (LLM) development. We address this gap by introducing an evaluation protocol for multimodal large language model (MLLM) benchmarking that explicitly accounts for two conditions: (1) human label agreement and (2) disagreement. We apply this protocol to two state-of-the-art MLLM families (Gemma 3, Qwen 2.5 VL) using non-aggregated human annotations from a social media content classification dataset. Across tasks, we find that larger models tend to perform best on high-agreement subsets, yet often underperform medium-sized models when human disagreement is high, indicating that parameter count alone does not determine sensitivity to ambiguity and subjectivity. These results show that benchmarks based solely on consensus labels can overstate model capabilities in such domains and that incorporating human label variation yields more realistic and robust assessments of MLLMs in content moderation pipelines.