A team led by Qi Zhang published a paper on 11 Nov 2026 showing that language models trained on images and text develop a hidden “vision axis” that they use during pure‑text reasoning. In the 7‑billion‑parameter model Chameleon‑7B, disabling this axis cut accuracy on the BIG‑Bench Hard navigation task by 27.6 percentage points while slightly boosting performance on symbolic boolean tests. The effect appears in specific layers and also helps with mental‑rotation and perspective‑taking tasks.
Why it matters
It shows that vision‑language training can boost spatial reasoning in text‑only AI, guiding future model design.
