Saved in:
| Main Authors: | , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2601.09954 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866918301026746368 |
|---|---|
| author | Alam, Nahid Murali, Leema Krishna Bharadwaj, Siddhant Liu, Patrick Chung, Timothy Sharma, Drishti A, Akshata Kiran, Kranthi Tam, Wesley Vegesna, Bala Krishna S |
| author_facet | Alam, Nahid Murali, Leema Krishna Bharadwaj, Siddhant Liu, Patrick Chung, Timothy Sharma, Drishti A, Akshata Kiran, Kranthi Tam, Wesley Vegesna, Bala Krishna S |
| contents | Vision-language models (VLMs) have advanced rapidly, but their ability to capture spatial relationships remains a blindspot. Current VLMs are typically built with contrastive language-image pretraining (CLIP) style image encoders. The training recipe often flattens images into 1D patch sequences, discarding the 2D structure necessary for spatial reasoning. We argue that this lack of spatial awareness is a missing dimension in VLM design and a bottleneck for applications requiring spatial grounding, such as robotics and embodied AI. To address this, we investigate (i) image encoders trained with alternative objectives and (ii) 2D positional encodings. Our experiments show that these architectural choices can lead to improved spatial reasoning on several benchmarks. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2601_09954 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | The Spatial Blindspot of Vision-Language Models Alam, Nahid Murali, Leema Krishna Bharadwaj, Siddhant Liu, Patrick Chung, Timothy Sharma, Drishti A, Akshata Kiran, Kranthi Tam, Wesley Vegesna, Bala Krishna S Computer Vision and Pattern Recognition Vision-language models (VLMs) have advanced rapidly, but their ability to capture spatial relationships remains a blindspot. Current VLMs are typically built with contrastive language-image pretraining (CLIP) style image encoders. The training recipe often flattens images into 1D patch sequences, discarding the 2D structure necessary for spatial reasoning. We argue that this lack of spatial awareness is a missing dimension in VLM design and a bottleneck for applications requiring spatial grounding, such as robotics and embodied AI. To address this, we investigate (i) image encoders trained with alternative objectives and (ii) 2D positional encodings. Our experiments show that these architectural choices can lead to improved spatial reasoning on several benchmarks. |
| title | The Spatial Blindspot of Vision-Language Models |
| topic | Computer Vision and Pattern Recognition |
| url | https://arxiv.org/abs/2601.09954 |