AI learns to predict the world without needing labeled training data
A new AI framework teaches itself to model environments by guessing the missing actions.
Deep Dive
Researchers introduced SWIRL, a framework where AI models learn to predict future states from past states by treating the connecting actions as a hidden, latent variable. It alternates between predicting the future and inferring the missing actions, training itself without costly action-labeled data. In tests on language and vision models, SWIRL improved performance by 14% to 28% on benchmarks for predicting physics, web navigation, and tool use.
Why It Matters
This reduces the need for expensive, hand-labeled data, making AI systems more efficient and capable of learning from observation alone.