arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2501.00569cs.CVcs.LG
收藏

探测视觉语言模型(VLMs)中的视觉语言先验

Probing Visual Language Priors in VLMs

University of Michigan(密歇根大学);LGAI Research(LGAI研究院)

Tiange Luo, Ang Cao, Gunhee Lee, Justin Johnson, Honglak Lee

AI总结:

为探究视觉语言模型(VLMs)过度依赖视觉语言先验的问题,本文提出ViLP基准,发现人类准确率接近完美而GPT-4仅66.17%,并提出自改进框架提升开源VLMs性能。

摘要:

尽管视觉语言模型(VLMs)近期取得进展,但它们可能过度依赖训练数据中存在的视觉语言先验,而非真实的视觉推理。为研究这一问题,我们引入ViLP基准,其包含通过图像生成模型合成的刻意分布外图像及分布外问答对。ViLP中的每个问题配有三个可能答案和三张对应图像:一张仅通过文本先验即可解决,另两张需视觉推理。尽管人类准确率接近完美,现代VLMs却表现不佳;例如,GPT-4在ViLP上仅得66.17%。为缓解此问题,我们提出一种自改进框架,模型生成新VQA数据,再应用像素级和语义损坏形成“好坏”图像对进行自训练,以促使VLMs更关注实际视觉输入。

英文摘要:

Despite recent advances in Vision-Language Models (VLMs), they may over-rely on visual language priors existing in their training data rather than true visual reasoning. To investigate this, we introduce ViLP, a benchmark featuring deliberately out-of-distribution images synthesized via image generation models and out-of-distribution Q&A pairs. Each question in ViLP is coupled with three potential answers and three corresponding images: one that can be resolved by text priors alone and two that demand visual reasoning. Although, humans achieve near-perfect accuracy, modern VLMs falter; for instance, GPT-4 achieves only 66.17% on ViLP. To alleviate this, we propose a self-improving framework in which models generate new VQA data, then apply pixel-level and semantic corruptions to form "good-bad" image pairs for self-training. Our training objectives compel VLMs to focus more on the actual visual inputs, and we demonstrate their effectiveness in boosting the performance of open-source VLMs, including LLaVA-v1.5 and Cambrian.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑