Research map / Architectures and efficiency / RNNs and normalization
Symbol-Equivariant Recurrent Reasoning Models
Richard Freinschlag, Timo Bertram, Erich Kobler, Andreas Mayr, Günter Klambauer
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add symbol-equivariant layers (symbol-dimension attention) to the recurrent reasoning model | SE-RRM vs TRM · 9x9 Sudoku | fully solved rate (%): 71.94 -> 93.73 | Improved |
| add symbol equivariance so unseen symbols can be embedded at test time | SE-RRM trained on 9x9 · 4x4 Sudoku zero-shot | fully solved rate (%): 0 -> 95.46 | Improved |
| apply SE-RRM to larger grids without training | SE-RRM trained on 9x9 · 16x16 and 25x25 Sudoku | fully solved rate (%): 0 FSR; GPA 51.95 and 31.49 | No clear effect |
| increase inference deep supervision steps from 1 to 128 | SE-RRM · 9x9 Sudoku | fully solved rate (%): 16.05 -> 98.84 | Improved |
| replace extensive color augmentation with only 8 dihedral augmentations using SE-RRM | SE-RRM · ARC-AGI-1 | pass@2 (%): TRM 44.6 -> SE-RRM 45.3 | No clear effect |
| use SE-RRM instead of TRM on ARC-AGI-2 | SE-RRM · ARC-AGI-2 | pass@2 (%): 7.8 -> 7.1 | No clear effect |
| use SE-RRM with symbol equivariance broken (separate symbol embeddings) instead of TRM | SE-RRM · Maze-hard | fully solved rate (%): 85.3 -> 88.8 | Improved |
| use SE-RRM instead of GPT-OSS-20B on 9x9 Sudoku | SE-RRM · 9x9 Sudoku | fully solved rate (%): 21.67 -> 93.73 | Improved |
Closest papers by meaning
- Causal language modeling can elicit search and reasoning capabilities on logic puzzles (NeurIPS 2024 · similarity 0.87)
- Boosting Inference with Guided Reasoning: Stochastic Exploration for Recursive Models (NeurIPS 2026 · similarity 0.86)
- Neural-Symbolic Recursive Machine for Systematic Generalization (ICLR 2024 · similarity 0.85)
- C-Voting: Confidence-Based Test-Time Voting without Explicit Energy Functions (ICLR 2026 · similarity 0.85)
- Verification Learning: Make Unsupervised Neuro-Symbolic System Feasible (ICML 2025 · similarity 0.85)
- Enigmata: Scaling Logical Reasoning in Large Language Models with Synthetic Verifiable Puzzles (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
