Symbol-Equivariant Recurrent Reasoning Models

2026-03-02 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Symbol-Equivariant Recurrent Reasoning Models (SE-RRMs) are introduced as a new architecture designed to explicitly enforce permutation equivariance in neural networks tackling reasoning problems like Sudoku and ARC-AGI. Unlike prior Recurrent Reasoning Models (RRMs) such as Hierarchical Reasoning Model (HRM) and Tiny Recursive Model (TRM), which rely on expensive data augmentation for symbol symmetries, SE-RRMs integrate symbol-equivariant layers to guarantee consistent solutions under symbol or color permutations. This architectural enhancement allows SE-RRMs to surpass existing RRMs on 9x9 Sudoku and generalize effectively from 9x9 training to smaller 4x4 and larger 16x16 and 25x25 instances, a capability prior RRMs lack. Furthermore, SE-RRMs achieve competitive performance on ARC-AGI-1 and ARC-AGI-2 with significantly less data augmentation and only 2 million parameters, demonstrating improved robustness and scalability through explicit symmetry encoding.

Key takeaway

For research scientists developing neural networks for structured reasoning problems, integrating symbol-equivariant layers into your models can significantly improve generalization across problem sizes and reduce reliance on extensive data augmentation. You should explore SE-RRMs to enhance robustness and scalability, particularly for tasks like Sudoku or ARC-AGI, where symbol permutations are inherent to the problem structure.

Key insights

Explicitly encoding symbol symmetry in neural networks enhances reasoning model robustness and scalability.

Principles

• Permutation equivariance improves generalization.
• Architectural symmetry reduces data augmentation needs.

Method

SE-RRMs enforce permutation equivariance via symbol-equivariant layers, guaranteeing identical solutions under symbol or color permutations, thus improving generalization across problem sizes.

In practice

• Apply SE-RRMs for structured reasoning tasks.
• Reduce data augmentation for symbol-symmetric problems.

Topics

• Recurrent Reasoning Models
• Symbol Equivariance
• Structured Reasoning
• ARC-AGI
• Sudoku

Code references

• ml-jku/SE-RRM

Best for: Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• ARC-AGI-2 Technical Report — Combining neural inference with structured priors and online adaptation significantly improves abstract reasoning on ARC-AGI.
• Grounding vs. Compositionality: On the Non-Complementarity of Reasoning in Neuro-Symbolic Systems — Symbol grounding alone is insufficient for compositional generalization; explicit reasoning objectives are required.
• Structured Abductive-Deductive-Inductive Reasoning for LLMs via Algebraic Invariants — A symbolic scaffold improves LLM logical reasoning by separating inference types and enforcing consistency via algebraic invariants.
• A homotopy-type-theoretic generalization of neurosymbolic inference — HoTT generalizes neurosymbolic inference, leveraging symmetries for improved shortcut-aware concept posteriors and calibration.
• Neural-Symbolic Logic Query Answering in Non-Euclidean Space — Hyperbolic GNNs with learned curvature effectively capture hierarchical logic query structures for knowledge graph reasoning.
• Structured Abductive-Deductive-Inductive Reasoning for LLMs via Algebraic Invariants — A symbolic reasoning scaffold improves LLM logical consistency by separating inference modes and enforcing algebraic invariants.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.