Figureure 2 · : Contrastive dataset construction overviewFigure 2: Contrastive dataset construction overview. In order to produce contrastive pairs, we mutate each base question four times with GPT-4o-mini and generate with and without-image responses to each variant. Then, for each pair of responses, we use cosine similarity and a regex annotator to identify mirage and non-mirage behavior. If a base question’s group contains at least one mirage and non-mirage pair, we store each pair’s with-image response to form a contrastive example.这张图概括 Mirage Probes 的整体方法流程。阅读时先看模块之间传递的训练信号，再看作者如何把目标拆成可优化的子问题。Figureure 1 · : Two distinct mirage mechanismsFigure 1: Two distinct mirage mechanisms. VLMs seem to exhibit two different kinds of mirage behavior, spurious images and textual biases, depicted here. Each row shows a mirage response to the same question with and without an image present. In the first example, the model achieves mirage behavior by building and referring to a false visual representation. Differently, in the second example, the model does not utilize visual information at all, relying solely on textual priors provided by the rich question distribution to reach the correct response.这张图来自论文 PDF 的结构化抽取。当前用于辅助理解 Mirage Probes 的方法或实验，请结合正文精读段落一起看。
