Research map / Vision-language and multimodal / Multimodal LLMs
Easy for Children, Hard for AI: The Limits of Multimodal LLMs in Early Childhood Learning
Jingping Liu, Xueyan Wu, Hanxuan Chen, Ziyan Liu, Zhangquan Chen, Ronghao Chen, Huacan Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Evaluates multimodal LLMs on early-childhood-style visual perception and reasoning tasks, finding them substantially weaker than the child-level skills they are meant to support.
Problem
Early childhood education is a critical developmental stage, but the ability of multimodal LLMs to support visual perception and reasoning for young learners is largely unexplored, and existing child-related AI focuses on language, emotion, or behavior.
Approach
Frames early-childhood visual perception and reasoning as an evaluation target for MLLMs, probing where current models fall short of tasks easy for children.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| LoRA fine-tune on ChildBench train (ablation) | Qwen2.5-VL-7B · ChildBench test | accuracy: 32.86 -> 52.01 | Improved |
| LoRA fine-tune on ChildBench train (ablation) | LLaMA-3.2-11B · ChildBench test | accuracy: 23.10 -> 25.35 | Improved |
| LoRA fine-tune on ChildBench train (ablation) | LLaVA-v1.5-7B · ChildBench test | accuracy: 17.68 -> 19.08 | Improved |
| LoRA fine-tune on ChildBench train (ablation) | DeepSeek-VL-7B · ChildBench test | accuracy: 17.50 -> 21.23 | Improved |
| LoRA fine-tune on ChildBench train (ablation) | MiniCPM-V-8B (Visual Discrimination) · ChildBench Visual Discrimination | accuracy: 25.64 -> 52.14 | Improved |
| add one in-context example (1-shot) (ablation) | GPT-4o · ChildBench (500 samples) | accuracy: 37.40 -> 29.20 | Hurt |
| add one in-context example (1-shot) (ablation) | Gemini-2.5-Pro · ChildBench (500 samples) | accuracy: 44.60 -> 45.80 | Improved |
| scale Qwen2.5-VL from 7B to 32B (ablation) | Qwen2.5-VL · ChildBench test | accuracy: 52.01 -> 55.00 | Improved |
| scale InternVL3.0 from 8B to 14B (ablation) | InternVL3.0 LoRA · ChildBench test | accuracy: 41.38 -> 40.40 | Mixed |
| scale InternVL3.0 from 8B to 38B (ablation) | InternVL3.0 LoRA · ChildBench test | accuracy: 41.38 -> 63.52 | Improved |
| merge multiple images into a single image (ablation) | Qwen2.5-VL-7B (LoRA) · ChildBench multi-image samples | accuracy: 43.48 -> 22.83 | Hurt |
| merge multiple images into a single image (ablation) | Gemini-2.5-Pro · ChildBench multi-image samples | accuracy: 19.40 -> 53.73 | Improved |
Surprising findings
- MLLMs have limits on early-childhood visual tasks that are easy for children
Limitations
- Only the abstract was available; experimental details unknown
Closest papers by meaning
- BabyVision: Visual Reasoning Beyond Language (ICML 2026 · similarity 0.91)
- DevBench: A multimodal developmental benchmark for language learning (NeurIPS 2024 · similarity 0.91)
- Children's Intelligence Tests Pose Challenges for MLLMs? KidGym: A 2D Grid-Based Reasoning Benchmark for MLLMs (ICLR 2026 · similarity 0.91)
- BabyVLM-V2: Toward Developmentally Grounded Pretraining and Benchmarking of Vision Foundation Models (CVPR 2026 · similarity 0.90)
- KiVA: Kid-inspired Visual Analogies for Testing Large Multimodal Models (ICLR 2025 · similarity 0.89)
- Core Knowledge Deficits in Multi-Modal Language Models (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
