Research map / Vision-language and multimodal / Multimodal LLMs

BabyVLM-V2: Toward Developmentally Grounded Pretraining and Benchmarking of Vision Foundation Models

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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.

Introduces BabyVLM-V2, a developmentally grounded pretraining set from SAYCam plus the DevCV Toolbox (ten vision tasks adapted from the NIH Baby Toolbox), and trains a compact from-scratch VLM that is compared against proprietary and open-source models.

Problem

Sample-efficient, developmentally plausible pretraining of vision-language foundation models lacks a comprehensive longitudinal infant-centric corpus and benchmarks grounded in established child cognitive tests.

Approach

Builds a longitudinal SAYCam-derived pretraining set (video-utterance, image-utterance, interleaved multi-turn), a compact LLaMA-1.1B + ViT-L VLM trained from scratch, and a benchmark of ten tasks adapted from NIH Baby Toolbox measures with naturalistic in-domain and Ego4D out-of-domain test sets.

Key results

• The from-scratch baby model reaches 63.9% overall on DevCV Toolbox in-domain, versus 74.6% for GPT-4o and 87.6% for GPT-5; it scores 99.7% on Who Has More (synthetic) vs. 79.3% for GPT-4o's Who Has More (naturalistic)-type comparison (see table).
• On Object Counting, the baby model (47.3%) outperforms GPT-4o (39.0%), and the synthetic-caption variant reaches 52.6%.
• Out-of-domain accuracy on the Ego4D sibling benchmark is 41.1% for the baby model, versus 31.8% random guess and 55.2% in-domain.

Surprising findings

• GPT-4o counts poorly, rarely going beyond 5 objects, while the small baby model outperforms it on counting given six or more objects.
• The baby model yields near-random results on the held-out tasks Looking While Listening and Subitizing.

Limitations

• Poor out-of-domain generalization to Ego4D.
• Near-random performance on unseen tasks.

Experimental setup

Datasets: SAYCam Ego4D MAB-CDI

Benchmarks: DevCV Toolbox NIH Baby Toolbox

Metrics: accuracy

Baselines: GPT-4o GPT-5 Gemini-2.5-pro Gemini-2.5-flash Qwen2.5-VL-7B Qwen2.5-VL-3B Qwen3-VL-30B-A3B LLaVA-OneVision-7B LLaVA-OneVision-0.5B InternVL3.5-1B

Builds on: BabyVLM-V1 LLaVA NIH Baby Toolbox DINOv2 X-CLIP

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 0.88)
• Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning (CVPR 2025 · similarity 0.87)
• Children's Intelligence Tests Pose Challenges for MLLMs? KidGym: A 2D Grid-Based Reasoning Benchmark for MLLMs (ICLR 2026 · similarity 0.87)
• What matters when building vision-language models? (NeurIPS 2024 · similarity 0.86)
• HoneyBee: Data Recipes for Vision-Language Reasoners (CVPR 2026 · similarity 0.86)