Effective knowledge condensation and reasoning models improve performance in knowledge-based VQA tasks.