Monday, 12 October 2026 · 5 min read
S-Space reveals an internal spatial coordinate system in multimodal language models
This paper discovers S-Space, a three-dimensional linear subspace encoding object locations in multimodal language models, enabling accurate spatial reasoning and diagnosis of viewpoint transformation failures, with causal evidence and generalization across tasks.
Based on the paper S-Space: Exploring Spatial Workspace in Multimodal Language Models by Zhiyuan Shi, Xiang Li, Xingzhuo Guo, Shangchen Miao and 7 others
Challenges in understanding spatial reasoning in multimodal language models
Multimodal language models (MLLMs) have shown promising spatial reasoning abilities useful for tasks like navigation and manipulation, but they still struggle with robust spatial reasoning, especially when viewpoint changes are involved.
Understanding why MLLMs fail spatial reasoning requires looking inside the model beyond output answers, as internal spatial representations and mechanisms remain poorly understood.
Inspired by human visual cognition's mental workspace concept, this paper investigates whether MLLMs have an internal spatial workspace encoding and manipulating spatial information to support reasoning.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Two key formulas from the paper, explained symbol by symbol
- Three figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short