Multimodal Halluciation Detection

Multimodal models can generate fluent and convincing responses that are nevertheless unsupported by the input, for example by referring to objects or details that are not present in an image. These hallucinations limit the reliability of vision-language models, especially in settings where users need to trust that generated outputs are grounded in the available evidence. Our research focuses on understanding why hallucinations occur, developing reliable methods to detect them, and designing mitigation strategies that reduce unsupported generations. We are particularly interested in signals that can identify hallucinations during generation, methods that distinguish reliable from unreliable model outputs, and inference-time approaches that improve factual grounding without substantially degrading the model's overall capabilities.