KidGym tests MLLMs on child-inspired tasks

This is a strong benchmark idea because it pushes MLLMs into a stateful setting where brittle reasoning shows up fast. The best part is interpretability: instead of one opaque score, KidGym separates memory, planning, counting, and compositional coordination.

• –The WISC-inspired structure turns evaluation into a capability profile that humans can actually reason about.
• –Randomized layouts and trajectory-based interaction reduce memorization and data leakage, so the signal should be closer to real generalization.
• –The backpack system, hint panel, and item indexing are a smart concession to current model limits without making the benchmark trivial.
• –The reported weak spots, abstract visual reasoning, numerical sensitivity, and multi-rule coordination, are exactly where many multimodal agents still fall apart.
• –As an extensible open benchmark, KidGym looks more useful for diagnosis and iteration than for leaderboard theater.