Hidden in Plain Sight: Occlusion Edge Blur as a Perceptual Blind Spot in Multimodal LLMs

Multimodal large language models (MLLMs) have demonstrated impressive reasoning capabilities across a range of visual tasks, yet emerging evidence suggests they struggle with fundamental perceptual processes. We investigate this gap using occlusion edge blur: a depth cue that humans exploit effortlessly but that cannot be solved by reasoning about a verbal description of the image by MLLMs. Our findings reveal a systematic bias in MLLMs likely driven by a verbal heuristic (“sharp edge means closer”) that overrides actual visual information in an image. That is, MLLMs can be easily ‘fooled’ on depth information by blurring the edge of an occluding object. This has implications for disinformation detection, robustness to adversarial attacks, and the design of new evaluation benchmarks.