Google and Cambridge taught AI to think with images instead of text

Scientists from Google and Cambridge presented a fundamentally new approach to artificial intelligence work, called “Visual Planning”. The feature of this technology is that the model thinks not with text, but with images. Which is much closer to natural human thinking, especially when solving spatial and mathematical tasks.

Researchers published an article titled “Visual Planning: Let’s Think Only with Images”, where they described the process of training a model to navigate mazes using only visual thinking, without textual reasoning. This approach mimics people’s ability to think with schemes and pictures when solving complex tasks. Link in the description.

Model training proceeded in two stages. First, it was shown many images of mazes and taught to predict any possible next step. For example, if in the picture an agent is located in cell B, the model should generate a new image where the agent moves to one of the available neighboring cells.

At the second stage researchers applied reinforcement learning. The model received positive reward for a correct step, zero — for incorrect, and negative — for an inadmissible action. Thus, gradually it learned to choose optimal paths through the maze. Relying only on visual images.

Results exceeded expectations! Visual Planning surpasses even such an advanced model as Gemini 2.5 Pro think by one and a half to two times in efficiency of solving tasks requiring spatial thinking. I am in pleasant shock!