An AI that teaches itself to see and draw — with no human labels

Most AI models that both understand pictures and create them need humans in the loop: people labeling images, ranking which answers are better, or a separate "judge" model scoring the work. This paper asks a sharper question: can the model improve at both jobs on its own, fed nothing but a pile of unlabeled photos?

The trick is splitting one model into three roles that quiz each other. The Proposer looks at an image and invents questions about it. The Solver answers those questions. The Generator creates new images. There's no answer key, so how does it know it's right? Consistency. If the Solver gives the same answer multiple times, that answer is probably solid — and that agreement becomes the reward signal. No outside grader needed.

One neat piece handles a weak spot: when answers stop agreeing cleanly, the simple "do they match?" check breaks down. So they also measure how confident the Solver is, word by word, as a backup difficulty gauge that keeps working when matching fails.

For drawing, the loop closes on itself: generate an image, then ask whether the Solver can answer questions about it correctly and re-describe it accurately. Better understanding makes for a stricter art critic, which trains a better artist.

Gains are real but modest — on one model, a 3.5-point bump on a knowledge test and image-generation accuracy from 82% to 85%. It improves things; it doesn't transform them.

Why you should care: Self-improvement without human labels means these gains came essentially for free — and the same recipe worked across three very different model designs.