The field of artificial intelligence is moving towards developing more accurate and reliable models, with a focus on reducing hallucinations. Hallucinations, which refer to the generation of factually incorrect text or images, pose a significant challenge in large language and vision models. Recent research has shown that finetuning models on high-quality factual information can help mitigate hallucinations, but the process of obtaining such data can be expensive and time-consuming. Moreover, training on correct but unfamiliar data may even lead to more downstream hallucinations. To address this issue, researchers are exploring alternative approaches, such as using model-generated data that models believe to be factual, or developing new architectures that can better align multimodal features. These innovative approaches have shown promising results in reducing hallucinations and improving the overall performance of AI models. Noteworthy papers in this area include: The Curious Case of Factuality Finetuning, which found that finetuning on model-generated data that models believe to be factual can lead to better factuality than finetuning on gold data. MCA-LLaVA, which proposed a new attention mechanism that integrates one-dimensional sequence order and two-dimensional spatial position of image tokens to mitigate hallucinations. Tracing Facts or just Copies, which investigated the role of attention heads in managing competing factual and counterfactual information. Mitigating Object Hallucinations via Sentence-Level Early Intervention, which introduced a framework that eliminates dependency on human annotations and reduces hallucinations by over 90%.