Researchers from the arXiv preprint investigate how vision-language models (VLMs) can learn to think and actively explore in three-dimensional space. The key design choice is to avoid providing the model with multi-view images, videos, or explicit 3D representations. Instead, the model's native visual encoding is kept intact, meaning it processes visual information in its original form.
By preserving the native visual encoding, the work asks whether VLMs can develop spatial understanding and exploration behavior directly from their standard visual input. The abstract does not disclose specific results or methods beyond this approach, so the full paper would be needed to assess the outcomes.