In the Experiments section of the LLaVA paper, the authors comprehensively evaluate the model's performance through two core experiments: multimodal chatbot conversational ability and the ScienceQA scientific question-answering task. For multimodal chatbot evaluation, using the constructed LLaVA-Bench benchmark (including COCO and In-the-Wild subsets), LLaVA demonstrated strong performance on instruction-following tasks, achieving relative scores of 85.1% and 67.3% respectively, significantly outperforming baseline models such as BLIP-2 and OpenFlamingo and showcasing robust vision-language alignment capability. On the ScienceQA multimodal reasoning benchmark, LLaVA alone achieved 90.92% accuracy, approaching the then-SOTA level, and when integrated with GPT-4 (using GPT-4 as a judge for decision-making), accuracy further improved to 92.53%, setting a new SOTA record. Ablation studies validated key design choices — such as using the penultimate layer features of the CLIP vision encoder, the chain-of-thought generation order, and the necessity of end-to-end instruction tuning — fully demonstrating the effectiveness of the visual instruction tuning strategy.