A new AI research paper argues machines should learn to explain the world, not just guess what comes next in it.
The paper introduces "Learning-to-Theorize," a training approach built around a system called NEO, short for Neural Theorizer. Instead of treating understanding as accurate prediction, which is how most current "world models" are trained and graded, NEO builds explicit, executable programs to explain raw, non-text observations. Those programs work like a compositional "language of thought," with learned building blocks that can be recombined to explain situations the system has not seen before. The researchers report this lets the model generalize through explanation, tying new observations back to the programs that generated them, rather than through pattern-matching alone.
That distinction matters because most world models today, including the video-generation and robotics-planning kind, are scored on how well they predict pixels or latent states, a proxy that says little about whether a system actually understands anything. A model that represents its knowledge as inspectable programs is at least theoretically easier to audit than one that represents it as inscrutable weights. It is also, notably, still an academic result running on curated experiments, not a deployed system.
Symbolic program induction has tempted AI researchers for decades and has just as often stalled once it meets messy, real-world data. NEO looks like the latest attempt to get machines to reason like scientists instead of just very good improvisers.