Overview

ARC-AGI tests whether systems can infer a task’s underlying rule from a few examples, usually by transforming colored grids in ARC-AGI-1 and ARC-AGI-2 or adapting to interactive environments in ARC-AGI-3. Approaches include program synthesis with domain-specific languages, neural transformers with test-time training, search over candidate rules, and agents that build executable world models. Systems can score highly on particular public evaluations, but performance does not reliably transfer across benchmark versions or establish robust generalization to new tasks. A central tension is how to measure genuine abstraction while guarding against shortcuts and accounting for interaction efficiency, compute cost, and the persistent gap from human performance on harder settings.

Latest papers

31 papers11 papers in the last four weeks, up 175% on the four weeks before. 0.1% of all new papers.

Mar 2, 2026

Symbol-Equivariant Recurrent Reasoning Models

Reasoning problems such as Sudoku and ARC-AGI remain challenging for neural networks. The structured problem solving architecture family of Recurrent Reasoning Models (RRMs), including Hierarchical Reasoning Model (HRM) and Tiny Recursive Model (TRM), offer a compact alternative to large language models, but currently handle symbol symmetries only implicitly via costly data augmentation. We introduce Symbol-Equivariant Recurrent Reasoning Models (SE-RRMs), which enforce permutation equivariance at the architectural level through symbol-equivariant layers, guaranteeing identical solutions under symbol or color permutations. SE-RRMs outperform prior RRMs on 9x9 Sudoku and generalize from just training on 9x9 to smaller 4x4 and larger 16x16 and 25x25 instances, to which existing RRMs cannot extrapolate. On ARC-AGI-1 and ARC-AGI-2, SE-RRMs achieve competitive performance with substantially less data augmentation and only 2 million parameters, demonstrating that explicitly encoding symmetry improves the robustness and scalability of neural reasoning. Code is available at https://github.com/ml-jku/SE-RRM.