Research

Paper

AI LLM March 20, 2026

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

Authors

Tomas Ruiz, Tanalp Agustoslu, Carsten Schwemmer

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.

Metadata

arXiv ID: 2603.19744
Provider: ARXIV
Primary Category: cs.CL
Published: 2026-03-20
Fetched: 2026-03-23 16:54

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.19744v1</id>\n    <title>Rethinking Ground Truth: A Case Study on Human Label Variation in MLLM Benchmarking</title>\n    <updated>2026-03-20T08:29:05Z</updated>\n    <link href='https://arxiv.org/abs/2603.19744v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.19744v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>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.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CL'/>\n    <published>2026-03-20T08:29:05Z</published>\n    <arxiv:comment>6 pages, 3 tables, 1 figure</arxiv:comment>\n    <arxiv:primary_category term='cs.CL'/>\n    <arxiv:journal_ref>2025 IEEE International Conference on Big Data (BigData), 2025</arxiv:journal_ref>\n    <author>\n      <name>Tomas Ruiz</name>\n    </author>\n    <author>\n      <name>Tanalp Agustoslu</name>\n    </author>\n    <author>\n      <name>Carsten Schwemmer</name>\n    </author>\n    <arxiv:doi>10.1109/BigData66926.2025.11401919</arxiv:doi>\n    <link href='https://doi.org/10.1109/BigData66926.2025.11401919' rel='related' title='doi'/>\n  </entry>"
}