引导式测验 · 中等 等级
视觉推理测验
测试图像感知与推理、图表比例和单位、受控和现实场景、小文本、语言先验和证据检查。
问题 1 的 8
A chart answer uses correct subtraction but reads the vertical axis incorrectly. Which stage failed?
预览每个问题8 问题
当您参加测验时,就会出现答案和解释。
- A chart answer uses correct subtraction but reads the vertical axis incorrectly. Which stage failed?
- Arithmetic after the values were extracted
- The format of the final sentence
- The choice of image file extension
- Perception of the visual evidence before inference
- A chart axis starts at 90; bars mark 95 and 100. What is their numerical difference?
- Ten units
- Five units
- Ninety-five units
- A difference cannot be read from a nonzero axis
- A model names every object in an image. What ability remains unproven?
- Recognition of object labels
- Production of a sentence
- Reliable counting, spatial comparison and relationship inference
- Detection of color differences
- Why pair controlled diagnostic scenes with realistic images in evaluation?
- A narrow skill test may not transfer to cluttered real material
- Simple scenes always prove deployment quality
- Real scenes make ground truth impossible
- Controlled data replace every task check
- A model misreads small chart text but reads large headings. Which input factor should be inspected?
- Only the model’s English vocabulary size
- Resolution and cropping of the relevant labels
- Only the chart’s publication date
- Only the answer’s punctuation
- 一张图片与对某个问题的最常见回答相矛盾。为什么这是一个有用的测试?
- 它消除了对源图像的需求
- 它保证所有视觉推理都会成功
- 它使图像标题成为法律记录
- 它探究语言先验是否会覆盖实际的视觉证据
- 模型对图表比较给出了详细的解释。审阅者应该做什么?
- 根据图表核对声明的标签、数值和关系
- 假设详细的推理证明了感知
- 仅验证语法
- 忽略底层图像
- 对于重要的图表使用,什么记录有助于重建答案?
- 仅最终的自然语言段落
- 仅置信度形容词
- 比较中使用的源图像和提取的数值
- 仅模型名称