A unified framework for early childhood image captioning via knowledge-grounded metrics and process-level reinforcement

Multimodal large language models (MLLMs) offer a promising solution for automating image captioning in early childhood education (ECE), where kindergartens caption large volumes of activity photographs daily for developmental portfolios and parent communication, a task that still relies on manual effort. Applying MLLMs to this domain faces three key challenges. First, existing evaluation metrics lack pedagogical domain knowledge and cannot detect terminology errors or material-level hallucinations. Second, without structured knowledge constraints, models frequently hallucinate non-existent educational materials, while child privacy regulations restrict deployment to locally executable architectures. Third, no reasoning-annotated training data exists for this domain, precluding chain-of-thought supervision that has proven effective elsewhere. We develop a knowledge graph-constrained evaluation metric that validates caption content against standardized kindergarten materials; propose a three-stage training pipeline integrating knowledge injection, structured chain-of-thought reasoning, and process-level reinforcement learning; and construct a domain-specific dataset with reasoning annotations aligned with national standards. Experiments demonstrate that our 7-billion-parameter (7B) model achieves performance comparable to counterparts at the 72-billion-parameter (72B) scale and to commercial application programming interfaces (APIs) across all evaluation dimensions while enabling fully local deployment for child privacy protection, and that our metric more than doubles the correlation of traditional metrics with expert ratings (Spearman 𝜌 = 0.701 vs. 0.308).