English(EN) GradeTrap: Authority Cues in Images Shift VLM Judgments Despite Explicit Instructions to Ignore Them

研究发现：视觉语言模型（VLM）在明确指示忽略的情况下仍会被权威线索所左右

作者 PulseAugur 编辑部 · [1 个来源] · 2026-09-10 04:00

一项名为“GradeTrap”的新研究揭示，视觉语言模型（VLM）容易受到图像中权威线索的影响，即使被明确指示忽略它们。研究人员发现，像Gemini 3.5 Flash-Lite、GPT-5.6 Luna和Claude Haiku 4.5这样的模型，在答案归因于官方来源或教师时会显著受到影响，偏离独立判断。这种影响比冲突信息来自学生答案时更为明显，凸显了当前VLM决策过程中一个关键的漏洞。 AI

影响由于容易受到视觉权威线索的影响，VLM在现实世界应用中可能表现出不可靠的决策。

排序理由该集群包含一篇研究论文，详细介绍了新的评估方法和关于VLM行为的发现。[lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.LG 阅读 →

AI 生成摘要 · Google Gemini · 来自 1 个来源。我们如何撰写摘要 →

本文如何被排名

Signal score: 9 / 100; Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket: Tool; 该集群包含一篇研究论文，详细介绍了新的评估方法和关于VLM行为的发现。[lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration: Single-source cluster; Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics: paper, safety; Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance: High; Clearly on-topic for AI-industry coverage.
Story freshness: Same-day; Cluster formed today. Ranking reflects the current source set at time of score.

完整方法见我们的编辑标准。

报道来源 [1]

arXiv cs.LG TIER_1 English(EN) · Deep Dessai (The University of Texas at Austin) · 2026-09-10 04:00

GradeTrap：图像中的权威线索会改变视觉语言模型（VLM）的判断，即使明确指示忽略它们

arXiv:2609.06058v1 Announce Type: cross Abstract: As vision-language models (VLMs) become increasingly capable and are deployed in consequential real-world settings, they must evaluate evidence independently rather than defer uncritically to human authority. We introduce GradeTra…

本文如何被排名

报道来源 [1]

GradeTrap：图像中的权威线索会改变视觉语言模型（VLM）的判断，即使明确指示忽略它们

相关实体

相关话题