| --- |
| license: mit |
| --- |
| |
| # FPRM — Fixed-Point Tiny Recursive Models |
|
|
| Checkpoints for **FPRM / FPTRM** (Fixed-Point Tiny Recursive Model), a weight-tied |
| iterative reasoner trained with a fixed-point solver and test-time compute scaling. |
|
|
| Each subfolder is one self-contained run (checkpoints + exact config + model source + |
| reproduction scripts). See each folder's `README.md` for the full recipe and results. |
|
|
| | folder | task | best test metric | notes | |
| |---|---|---|---| |
| | [`maze/`](maze) | Maze-Hard 30×30 | **87.0%** exact-match | single-z, conv1d/k4, non-augmented; eval @ max_iter 35k, decay 0.996/pat 10 | |
| | [`sudoku/`](sudoku) | Sudoku-Extreme | **94.2%** exact-match | single-z, conv2d/k3, norm-placement=none; eval @ max_iter 35k, decay 0.997/pat 10 | |
| | [`arc1/`](arc1) | ARC-1-concept (aug-1000) | **47.5%** pass@2 | single-z, conv1d/k4, norm-placement=none; eval @ max_iter 1000 | |
| | [`arc2/`](arc2) | ARC-2-concept (aug-1000) | **6.2%** pass@2 | single-z, conv1d/k4, norm-placement=none; eval @ max_iter 1000 | |
|
|
| Checkpoints are saved every 5000 epochs as `step_<N>` (EMA-averaged eval weights, the ones |
| to load) and `step_<N>_train_state.pt` (full training state for resuming). The `arc1/` and |
| `arc2/` folders ship eval checkpoints only (no `train_state`). |
|
|
