KAIST's 'Neural Theorizer' invents small programs to explain the world

Researchers at KAIST, the Korea Advanced Institute of Science and Technology, have proposed a new way to train AI: let a model watch raw, unlabeled video of falling blocks, rolling balls, and shifting pendulums, and have it write small, executable programs to explain what it sees. The result, the Neural Theorizer (NEO), reframes "understanding" less as accurate next-frame prediction and more as reverse-engineering the mechanism that produced the change. The paper was accepted as an ICML 2026 Oral, in the top 0.7% of 23,918 submissions per the KAIST MLML project page.

The work introduces a new learning paradigm the authors call Learning-to-Theorize (L2T). In NEO, a "theory" is not a vague intuition. It is an executable, compositional program whose learned primitives can be recombined to explain phenomena the model has never seen. A probabilistic neural model induces these latent programs as a kind of learned "Language of Thought," and a shared transition model executes them against observation. The bet is simple: if a model can write the rule, the rule is what generalizes. The training loop is built around that, inducing the program first and using it to predict later.

World models, the systems that learn a compressed representation of how an environment changes, are typically judged on how well they predict the next frame. L2T instead asks what rule produced the change, and the transition model is what makes that question answerable: a learned simulator that takes a candidate program and rolls it forward, so the model can score theories by how well their execution matches the observed data.

The framing is deliberately drawn from developmental cognitive science. The authors lean on a long tradition: children, the argument goes, build internal theories of the world before they have mature language, and those theories do the work of explanation. NEO is positioned as the machine-learning analogue, induce the theory first and defer the words. The analogy is the authors' own framing rather than field consensus, and the paper treats "understanding" as contested rather than achieved. That posture is part of what makes the work worth reading, not a caveat to apologize for.

The empirical work sits behind the v2 arXiv HTML and is still a preprint. ICML 2026 Oral is a review indicator, not a camera-ready stamp, and the project page also notes a CompLearn 2026 Best Paper award. First author Gyubin Lee submitted v1 on 5 May 2026 and v2 on 4 June 2026. Anyone weighing the result should anchor on the paper's reported experiments, not on aggregator framing. The Mirage News re-report of the release was Cloudflare-blocked during source hydration and adds no mechanism detail beyond the title; the primary signal is the arXiv preprint plus the KAIST MLML publications page.

A model that recovers a hidden rule from observation makes a different kind of claim than a model that just predicts better. The test is whether the learned representation looks like a mechanism: a few composable pieces that can be recombined and still hold. If the answer is yes, the work is doing mechanism inference. If the answer is no, it is better prediction wearing theory's costume.

The next time a headline says an AI "now understands the world," that is the question worth asking. The KAIST team at least has a paper that tries to answer it.