Open-ended Association Reasoning Benchmark for MLLMs

Overview

MM-OPERA explores open-ended association reasoning in multimodal large language models. My work focused on dataset quality and the reliability of both human and model-based evaluation.

My contributions

• Built a multi-stage quality-assurance pipeline across 11,497 instances.
• Independently led human baseline evaluations covering 921 tasks.
• Evaluated LLM-as-a-Judge reliability, including verbosity and position bias.

Areas & tools

• MLLMs
• Benchmark
• Human evaluation