Prompt Sensitivity in Vision-Language Grounding: How Small Changes in Wording Affect Object Detection
For practitioners using vision-language models for object detection, this reveals a fundamental reliability issue in grounding that is not resolved by prompt ensembling.
The paper shows that semantically equivalent prompts (e.g., 'a person', 'a human') cause vision-language grounding models to select different object instances, with mean instability of 2.11 distinct selections across six prompts. Text embedding proximity explains only 34% of this disagreement.
Vision-language models enable open-vocabulary object grounding through natural language queries, under the implicit assumption that semantically equivalent descriptions yield consistent outputs. We examine this assumption using a controlled pipeline combining DETR for object proposals with CLIP for language-conditioned selection on 263 COCO val2017 images. We find that overlapping prompts such as "a person," "a human," and "a pedestrian" frequently select different instances, with mean instability of 2.11 distinct selections across six prompts. PCA analysis shows this variability is structured and directional, not random. Prompt ensembling does not improve quality and often shifts selections toward generic regions. We further show that text embedding proximity explains only 34% of grounding disagreement (r = -0.58), confirming that instability arises from the argmax selection mechanism rather than text-level distances alone.