LEVANTE-bench: Multi-Scale Comparison of VLMs to Children Using Cognitive Tasks (or, "Is Your VLM Smarter Than a 5th Grader?")

In three linesLEVANTE-bench benchmarks vision-language models against children aged 5-12 (N=1547) on cognitive tasks across 3 countries. Larger models show better overall alignment, but smaller models match younger children's error patterns better. VLMs struggle on matrix reasoning and mental rotation tasks.

Summary generated by Claude — human-verified