Researchers from Peking University, Tsinghua University, Peng Cheng Laboratory, Alibaba DAMO Academy, and Lehigh University have introduced LLaVA-o1, a novel visual language model (VLM) capable of spontaneous and systematic reasoning. LLaVA-o1 distinguishes itself by conducting autonomous multistage reasoning through four sequential stages: summarization, visual interpretation, logical reasoning, and conclusion generation. This innovative approach contrasts with traditional chain-of-thought prompting methods. The model, which is fine-tuned from the Llama-3.2-11B-Vision-Instruct, has demonstrated superior performance across six multimodal reasoning benchmarks, outperforming larger models such as Gemini-1.5-pro, GPT-4o-mini, and even some closed-source alternatives. LLaVA-o1 is trained on 100,000 samples and employs a stage-level beam search to generate and select optimal answers at each reasoning stage.