A new research paper explores the vulnerabilities of Vision-Language Models (VLMs) to subtle changes in object descriptions, specifically focusing on landmark synonyms. The study reveals that VLMs can be easily misled by synonymous words, impacting their ability to complete tasks accurately. Researchers found that while synonyms can direct the VLA to the correct object, they significantly reduce task completion rates compared to original instructions. Interventions at various points in the VLA pipeline, including the connector between the vision-language model and action generator, were tested to improve robustness, with fine-tuning on diverse landmark words showing the most promise. AI
IMPACT Highlights a critical area for improving the robustness and reliability of vision-language models in real-world applications.
RANK_REASON Academic paper detailing a specific vulnerability in VLMs. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- Connected Papers
- DagsHub
- Gotit.pub
- GR00T-N1.7
- Hugging Face
- LIBERO-Spatial
- Litmaps
- ScienceCast
- scite Smart Citations
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →