# Multimodal training gives text‑only models hidden visual reasoning abilities

> Researchers show that vision‑language training reshapes how language models solve spatial tasks, even when only text is fed.

Oossa · 2026-11-13 · https://oossa.com/en/multimodal-training-gives-text-only-models-hidden-visual-reasoning-abilities

Event date: 2026-11-11

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.

## The facts

- Chameleon‑7B loses 27.6 pp accuracy on BIG‑Bench Hard navigate when its vision axis is ablated
- Paper submitted 11 Nov 2026 (Wed, 11 Nov 2026 02:44:10 UTC)

## Why it matters

It shows that vision‑language training can boost spatial reasoning in text‑only AI, guiding future model design.

## Sources & references

1. [Emergent Vision-Shaped Thinking in Text-Only Reasoning through Multimodal Training](https://arxiv.org/abs/2611.14942) – arXiv, 2026-11-13

Last updated: 2026-11-13