Papers11:00Chinese content失败引导的图像自增强：FISA提升多模态大模型自我改进能力这篇讲FISA，拿模型答错的题来造新训练图，比瞎做增强更准，还能和文本增强一起用。#FISA#multimodal LLMs#recursive self-improvement#visual question answeringaarXiv cs.AI@Chunyang Jiang et al. (12)SourceRead laterReadWorth followingUsefulFollow FISA