alphaXiv

This is a Tiny Recursive Model (TRM) fine-tuned for solving maze navigation tasks. The model implements recursive reasoning to find paths in 30x30 grid mazes. - Developed by: alphaXiv - Model type: TRM-Attention - Language(s) (NLP): N/A (grid-based reasoning) - License: MIT - Finetuned from model: Custom TRM architecture This model is designed to solve maze pathfinding problems by predicting the correct sequence of moves to navigate from start to goal in grid-based mazes. Not intended for general NLP tasks, image classification, or other domains outside maze solving. - Trained only on synthetic maze data - May not generalize to mazes of different sizes or complexities - Performance may degrade on mazes with unusual patterns The model was trained on a dataset of 30x30 grid mazes with hard difficulty levels. The dataset includes: - Start and goal positions - Wall configurations - Correct path sequences | Metric | Claimed | Achieved | |--------|---------|----------| | Exact Accuracy | 85.3% | 83.67% ± 2.28% |

This is a Tiny Recursive Model (TRM) fine-tuned for solving Abstract Reasoning Challenge (ARC-AGI) tasks. The model performs abstract reasoning to predict output grids from input grids. - Developed by: alphaXiv - Model type: TRM-Attention - Language(s) (NLP): N/A (grid-based reasoning) - License: MIT - Finetuned from model: Custom TRM architecture This model is designed to solve ARC-AGI tasks by predicting the correct output grid transformation based on input grid patterns. Not intended for general NLP tasks, image generation, or other reasoning domains. - Trained only on ARC-AGI training and evaluation sets - May not generalize to novel abstract reasoning tasks - Performance limited by training data diversity The model was trained on the ARC-AGI dataset, which includes: - Input-output grid pairs - Various transformation patterns - Training and evaluation splits | Metric | Claimed | Achieved | |--------|---------|----------| | Pass@2 | 44.6% | 43.00% ± 0.16% |

This is a Tiny Recursive Model (TRM) fine-tuned for solving Sudoku puzzles. The model uses recursive reasoning to fill in missing numbers in Sudoku grids. - Developed by: alphaXiv - Model type: TRM-MLP - Language(s) (NLP): N/A (grid-based reasoning) - License: MIT - Finetuned from model: Custom TRM architecture This model is designed to solve Sudoku puzzles by predicting the correct numbers for empty cells in standard 9x9 Sudoku grids. Not intended for general NLP tasks, image processing, or other puzzle types. - Trained only on standard 9x9 Sudoku puzzles - May not handle non-standard Sudoku variants - Performance depends on puzzle difficulty The model was trained on a dataset of Sudoku puzzles with extreme difficulty levels. The dataset includes: - Partially filled 9x9 grids - Correct solutions - Difficulty ratings | Variant | Metric | Claimed | Achieved | |---------|--------|---------|----------| | TRM-MLP | Accuracy | 87.4% | 79.37% ± 0.12% | | TRM-Attention | Accuracy | 74.7% | 73.66% ± 0.13% |