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

A new benchmark called LEVANTE-bench systematically compares vision-language models against the cognitive abilities of children aged 5 to 12 across multiple countries, finding that even the strongest models struggle with matrix reasoning and mental rotation tasks [1]. The framework, detailed in a paper submitted June 3, draws on tasks and data from the Learning Variability Network, an open-source project measuring children's cognition across languages and cultures [1]. Researchers evaluated VLMs on six tasks, comparing model outputs with data from 1,547 children in three countries [1]. The study assessed overall accuracy, task- and item-level alignment with children, and how closely models matched children's trial-level error distributions [1]. Alignment proved heterogeneous across scales. At the task and item levels, more capable models aligned better with humans [1]. However, match to human error distributions varied widely, and for several tasks smaller models matched younger children's errors more closely [1]. The paper notes that current VLM architectures align only partially with the cognitive abilities of children [1]. The benchmark arrives amid broader efforts to measure how artificial systems handle reasoning tasks that humans find intuitive. A separate study on long-video understanding found a strong positive linear correlation between a VLM's performance on logic reasoning and long-video understanding benchmarks, suggesting that agentic capability scaling may offer a path toward more robust multimodal comprehension [2]. That work introduced a hierarchical graph memory architecture that constrained the reasoning context window to 2 percent of full-context ingestion while delivering a 12.5 point absolute accuracy gain [2]. LEVANTE-bench's multi-scale comparison approach departs from typical AI benchmarks that report only aggregate accuracy. By examining error distributions across age groups, the benchmark reveals where model failures mirror developmental patterns and where they diverge. The finding that even the best-performing VLMs struggled on matrix reasoning and mental rotation tasks underscores a gap between statistical pattern matching and the structured spatial reasoning that children develop [1]. The work was led by Alvin Wei Ming Tan and collaborators [1]. The paper is available on arXiv alongside code and data links [1].