Science Domain Visual Understanding

via “science-domain-visual-understanding”

Open multimodal model for visual reasoning.

Unique: Achieves 92.53% Science QA accuracy through general instruction-tuning without explicit science-domain fine-tuning, suggesting the GPT-4-generated reasoning samples capture sufficient scientific reasoning patterns; this emergent domain capability differs from models requiring explicit domain adaptation

vs others: Outperforms general-purpose vision-language models on Science QA without domain-specific training because its instruction-tuning dataset includes diverse reasoning patterns that generalize to scientific domains

via “visual concept explanation with diagrams”