Search papers, labs, and topics across Lattice.
The paper introduces Symbol-Equivariant Recurrent Reasoning Models (SE-RRMs), a novel architecture that enforces permutation equivariance in recurrent reasoning models using symbol-equivariant layers. This explicit encoding of symmetry allows SE-RRMs to generalize to different problem sizes (e.g., Sudoku grids) without retraining and reduces the need for extensive data augmentation. Experiments show that SE-RRMs outperform previous RRMs on Sudoku and achieve competitive results on ARC-AGI with significantly less data augmentation and a smaller model size.
By baking in symbol equivariance, SE-RRMs solve Sudoku and ARC-AGI problems with far less data augmentation and fewer parameters than previous 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.
