Researchers find that enriching image captions with semantic knowledge, rather than adding VQA tasks, is the key to scaling multimodal AI models more effectively.

VQA (Visual Question Answering) training contributes minimal new information beyond image captions and can be reconstructed from captions with negligible performance loss

Knowledge density in training data, not task format diversity, is identified as the primary bottleneck limiting multimodal LLM scaling performance

Structured caption enrichment and cross-modal knowledge injection produce consistent improvements across multimodal and downstream benchmarks

Performance gains correlate more strongly with semantic coverage than with increasing model size or task diversity alone