A Systematic Evaluation of Positional Bias in Multi-Video Summarization with MLLMs

In three linesSystematic study of positional bias in multi-video summarization with MLLMs. Benchmark on ActivityNet and News videos (2-4 inputs). Evaluation of 9 models (open-source and proprietary) using three metrics: Coverage, Directional Positional Bias, Middle-Edge Gap. Finding: positional effects are domain- and model-dependent; increasing visual budget does not uniformly remove imbalance.

Summary generated by Claude — human-verified