Imagine asking an AI assistant about a photo on your phone and getting a completely fabricated description. This frustrating phenomenon, called "hallucination," poses a significant challenge for advanced AI models that process both images and text (Vision-Language Models or VLMs). These powerful models are designed to understand and reason about visual scenes, generating captions, answering questions, and even following complex instructions. However, they sometimes generate nonsensical or factually incorrect outputs, limiting their real-world applications. Researchers at SRI International have introduced Pelican, a novel framework to combat these AI hallucinations. Pelican acts as a fact-checker for VLMs, scrutinizing their claims about images through a process called claim decomposition and verification. Think of it like breaking down a complex argument into smaller, verifiable statements. Pelican dissects a VLM's response into a chain of sub-claims, each represented by a question. For example, if a VLM claims, "A man wearing glasses is riding a motorcycle," Pelican might generate sub-claims like, "Is there a man?", "Is he wearing glasses?", "Is he riding something?", "Is it a motorcycle?". To answer these questions, Pelican ingeniously uses a 'Program of Thought' approach. This involves prompting another large language model (LLM) to create Python code that utilizes readily available visual tools like object detectors and Visual Question Answering (VQA) systems. The Python code effectively acts as a bridge between Pelican’s reasoning and the concrete visual evidence present in the image. Pelican's innovations include introducing intermediate variables for accurate object grounding, sharing calculations between sub-questions for efficiency and consistency, and reasoning in both natural language and computational methods. For instance, it might identify the "man" using a bounding box and then verify whether or not he is wearing “glasses” within that region. Tests using various benchmarks like MMHal-Bench, GAVIE, and MME showed Pelican significantly reduces hallucinations and improves accuracy across different VLMs. Notably, it achieved up to a 32% reduction in hallucination rates on MMHal-Bench and a substantial 27% drop compared to other hallucination mitigation approaches. Pelican's innovative approach marks a critical step towards creating more trustworthy and reliable VLMs. Its ability to systematically identify and correct hallucinations strengthens these models' potential for real-world applications, from assisting visually impaired users to automating complex analyses of images and videos. Although the research highlights some limitations like the brittleness of the tools, Pelican points towards a brighter future for multimodal AI models that can reliably see and understand the world around us.