Apple researchers have developed RubiCap, a new AI framework that generates more accurate image captions using models up to 36 times smaller than competing systems. The breakthrough addresses a fundamental challenge in dense image captioning, where AI must describe multiple elements within a single image with fine-grained detail rather than providing a single summary.

Traditional approaches struggle with either the cost of human annotation or the limitations of synthetic data. RubiCap's solution uses reinforcement learning with a novel rubric-guided approach. The system generates multiple caption options using large vision-language models like Gemini 2.5 Pro and GPT-5, then uses these as benchmarks to train smaller models. A judging component scores captions against specific criteria, providing structured feedback without requiring a single 'correct' answer.

The results are striking: RubiCap's 3-billion-parameter model outperformed Qwen2.5-VL-7B-Instruct, while the 7-billion-parameter version matched or exceeded models with up to 72 billion parameters. In blind evaluations, RubiCap-7B achieved the highest proportion of top-ranked assignments among all tested models, including industry-leading 72B and 32B systems.