mair-lab/Cosmic

---
configs:
- config_name: anchor_recognition
data_files:
- split: test
path: anchor_recognition/test-00000.parquet
- config_name: cognitive_mapping
data_files:
- split: test
path: cognitive_mapping/test-00000.parquet
- config_name: global_counting
data_files:
- split: test
path: global_counting/test-00000.parquet
- config_name: relative_direction
data_files:
- split: test
path: relative_direction/test-00000.parquet
- config_name: relative_distance
data_files:
- split: test
path: relative_distance/test-00000.parquet
---

<h1 align="center">Communicating about Space: Language-Mediated Spatial Integration Across Partial Views</h1>

<p align="center">
<a href="https://arxiv.org/abs/2603.27183"><img alt="arXiv" src="https://img.shields.io/badge/arXiv-Comm--About--Space-red?logo=arxiv" height="20"/> </a>
<a href="https://github.com/ankursikarwar/Cosmic"><img src="https://img.shields.io/badge/github-Comm--About--Space-blue?logo=github" alt="Github"/> </a>
</p>

<div align="justify">

Humans routinely transform local, viewpoint-dependent observations into shared spatial models through language. COSMIC asks whether MLLMs can do the same. The benchmark places two static agents in the same indoor scene from different egocentric viewpoints. The agents must communicate exclusively through natural language to jointly solve a spatial QA task.

</div>

## Usage

```python
from datasets import load_dataset

ds_anchor_recognition = load_dataset("mair-lab/Cosmic", name="anchor_recognition", split="test")
ds_cognitive_mapping = load_dataset("mair-lab/Cosmic", name="cognitive_mapping", split="test")
ds_global_counting = load_dataset("mair-lab/Cosmic", name="global_counting", split="test")
ds_relative_direction = load_dataset("mair-lab/Cosmic", name="relative_direction", split="test")
ds_relative_distance = load_dataset("mair-lab/Cosmic", name="relative_distance", split="test")

```

## Citation

```bibtex
@misc{sikarwar2026communicatingspacelanguagemediatedspatial,
title={Communicating about Space: Language-Mediated Spatial Integration Across Partial Views},
author={Ankur Sikarwar and Debangan Mishra and Sudarshan Nikhil and Ponnurangam Kumaraguru and Aishwarya Agrawal},
year={2026},
eprint={2603.27183},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2603.27183},
}
```