Hidden 3‑D Coordinates Inside GPT‑4V? S‑Space Reveals the Secret
S‑Space Turns Existing Multimodal LLMs into 3‑D Workspaces Headline Meta AI’s October 2026 “S‑Space” release exposes a hidden linear coordinate system inside vision‑language models, letting developers pull 3‑D positions out of any visual token without extra spatial training. The Lead “We discovered a linear sub‑space that already encodes x, y and depth inside GPT‑4V‑style backbones.” – Lead…
Meta AI unveiled S-Space, a technique that converts vision-language models like GPT-4V into 3D workspaces capable of extracting coordinates from visual tokens. The paper, published on arXiv in October 2026, revealed that existing multimodal models already encode spatial information in a hidden linear coordinate system. S-Space offers a Python SDK and demos that turn any vision-language transformer into a spatial workspace, without the need for additional spatial training.
The technique involves discovering a linear sub-space that encodes x, y, and depth coordinates based on the model's activations, and an adapter that projects token embeddings onto orthogonal basis vectors representing horizontal, vertical, and depth directions. Benchmarks showed significant improvements in multi-step spatial inference and continuous distance estimation compared to the same model without the S-Space adapter.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
