Many reasoning tasks demand deducing latent laws from few demonstrations to solve unseen problems. It remains unclear whether compact neural architectures, through appropriate inductive biases, can rival the reasoning abilities that large models achieve through scale.
We introduce Interactive Endomorphic Reasoning Models (IERM), a compact neural architecture that explicitly separates two computational roles: the induction of latent programs from demonstrations and their execution during iterative reasoning.
IERM organizes computation into three interacting representation spaces: a support-induced program space encoding transformation laws inferred from examples, a compact reasoning space serving as a working memory, and a solution space responsible for constructing candidate outputs. These components interact through recursive cross-attention updates, enabling progressive refinement of solutions while maintaining compact internal representations.
Despite using very small models, IERM achieves around 12% solved tasks on ARC-AGI-1 and over 63% on Sudoku Extreme using only 2–4M parameters.
These results suggest that reasoning ability does not necessarily require massive model scale, but can emerge from appropriate architectural inductive biases and recursive computation.