Rethinking Ground Truth: A Case Study on Human Label Variation in MLLM Benchmarking
This addresses the issue of unrealistic assessments in MLLM benchmarking for content moderation, though it is incremental as it applies an existing concept to a new domain.
The study tackled the problem of human label variation in multimodal large language model (MLLM) benchmarking by introducing an evaluation protocol that accounts for both agreement and disagreement among annotators, finding that larger models perform best on high-agreement subsets but underperform medium-sized models when disagreement is high, indicating benchmarks based on consensus labels can overstate capabilities.
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.