Mechanisms of Prompt-Induced Hallucination in Vision–Language Models
Автор: AI Papers Podcast Daily
Загружено: 2026-01-18
Просмотров: 9
Vision-Language Models (VLMs) often suffer from *prompt-induced hallucinations (PIH)**, which occur when a model trusts a written instruction more than the actual image it is looking at,. For instance, if a user asks a model to describe four flowers when only three are present, the model will often **hallucinate* the extra flower to match the text,. This behavior is most common when images contain more than four objects, as the model’s *visual confidence* decreases and it begins to rely more on the prompt,. Researchers found that a small group of *attention heads* within the model are responsible for this copying behavior, and by "turning off" (ablating) these heads, they reduced hallucinations by at least 40% without any extra training,. This discovery allows models to focus more on *visual evidence* and less on incorrect text, making them much more accurate in tasks ranging from counting to identifying colors,,.
https://arxiv.org/pdf/2601.05201
https://github.com/michalg04/prompt-i...
Доступные форматы для скачивания:
Скачать видео mp4
-
Информация по загрузке:
