dorsal/arxiv
View SchemaThe Spatial Blindspot of Vision-Language Models
| Authors | Nahid Alam, Leema Krishna Murali, Siddhant Bharadwaj, Patrick Liu, Timothy Chung, Drishti Sharma, Akshata A, Kranthi Kiran, Wesley Tam, Bala Krishna S Vegesna |
|---|---|
| Categories | |
| ArXiv ID | 2601.09954vv1 |
| URL | https://arxiv.org/abs/2601.09954 |
| License | http://creativecommons.org/licenses/by/4.0/ |
Abstract
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.
{
"annotation_id": "385a9b0b-2a30-4a03-aaf8-257ad9d8e7ee",
"date_created": "2026-02-17T05:53:24.373000Z",
"date_modified": "2026-02-17T05:53:24.373000Z",
"file_hash": "7a0272ca983568bb2339302aad7499a795f73b847488679355a64e133cec1b44",
"private": false,
"record": {
"abstract": "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.",
"arxiv_id": "2601.09954",
"authors": [
"Nahid Alam",
"Leema Krishna Murali",
"Siddhant Bharadwaj",
"Patrick Liu",
"Timothy Chung",
"Drishti Sharma",
"Akshata A",
"Kranthi Kiran",
"Wesley Tam",
"Bala Krishna S Vegesna"
],
"categories": [
"cs.CV"
],
"license": "http://creativecommons.org/licenses/by/4.0/",
"title": "The Spatial Blindspot of Vision-Language Models",
"url": "https://arxiv.org/abs/2601.09954",
"version": "v1"
},
"schema_id": "dorsal/arxiv",
"source": {
"execution_id": "934ede70-f2fd-4e1a-bf70-c3fe98589cd4",
"id": "arXiv Dataset",
"type": "Model",
"variant": "snapshot-2026-01-17",
"version": "0.1.0"
},
"user_id": 1000002
}