Superhuman Induction: Autonomous Algorithm Discovery in ARC-AGI via Latent Hypothesis Selection

The Abstraction and Reasoning Corpus (ARC-AGI) is a benchmark designed to measure the human-like ability to induce general-purpose rules from sparse data. While Transformer-based Large Language Models (LLMs) excel at template-matching, they consistently fail at the "core knowledge" tasks of ARC that require symbolic grounding and structural induction. We present a Structural Induction Agent based on Active Inference that treats algorithm discovery as a model selection problem. By navigating a discrete hypothesis space of architectural primitives, the agent autonomously identifies the latent generative rule of a puzzle from as few as three examples. Our system achieved 100% accuracy on complex multi-step transformations. This research demonstrates that general intelligence can be formalized as the optimal minimization of informational surprise across algorithmic hypotheses.