Prompt Sensitivity in Vision-Language Grounding: How Small Changes in Wording Affect Object Detection

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Deka, Dawar Jyoti, Sethi, Amit, Ali, Syed Mohammad
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866917418158260224
author Deka, Dawar Jyoti
Sethi, Amit
Ali, Syed Mohammad
author_facet Deka, Dawar Jyoti
Sethi, Amit
Ali, Syed Mohammad
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2604_17126
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Prompt Sensitivity in Vision-Language Grounding: How Small Changes in Wording Affect Object Detection
Deka, Dawar Jyoti
Sethi, Amit
Ali, Syed Mohammad
Computer Vision and Pattern Recognition
I.4.8; I.2.10; I.5.4
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.
title Prompt Sensitivity in Vision-Language Grounding: How Small Changes in Wording Affect Object Detection
topic Computer Vision and Pattern Recognition
I.4.8; I.2.10; I.5.4
url https://arxiv.org/abs/2604.17126