SFPRM checkpoints

Checkpoints of SFPRM (Spherical Fixed-Point Reasoners), the spherical variant of FPRM, and of the baselines it is compared with: FPRM, TRM, HRM and EqR. Code, configs, datasets and evaluation commands: github.com/arberzela/sfprm.

Contents

One folder per checkpoint, laid out as pretrain.py writes it:

File	Content
step_<N>	The weights the run evaluated at step N (EMA weights for every run except HRM). Evaluate this file.
step_<N>_train_state.pt	The resume bundle: live weights, optimizer state, EMA shadow and data-loader position. Not included for sfprm-arc-step518065.
all_config.yaml	The fully resolved training config.
PROVENANCE.md, *.py	Provenance notes and the model and loss source used by the run.

Checkpoint	Model	Task	Result
sfprm-sudoku-step97656	SFPRM	Sudoku-Extreme	96.58% exact accuracy (full test set, 422,786 puzzles)
sfprm-maze-step78120	SFPRM	Maze-Hard 30×30	87.9% exact accuracy (1000 mazes, in-run)
fprm-maze-step78120	FPRM	Maze-Hard 30×30	83.9% exact accuracy (1000 mazes, in-run)
sfprm-arc-step518065	SFPRM	ARC-AGI-1	41.25% / 46.62% pass@1 / pass@2 (in-run, pooled)
sfprm-miniarc-step68620	SFPRM	Mini-ARC	59.1% / 59.7% pass@1 / pass@2
fprm-miniarc-step68620	FPRM	Mini-ARC	55.0% / 62.4%
hrm-miniarc-step68620	HRM	Mini-ARC	51.7% / 58.4%
eqr-miniarc-step68620	EqR	Mini-ARC	40.3% / 49.7%
trm-miniarc-step68620	TRM	Mini-ARC	36.9% / 45.0%

Mini-ARC scores are single-checkpoint votes at a rollout cap of 200 (arch.max_iter_eval=200 for SFPRM/FPRM, arch.halt_max_steps=200 for TRM/HRM/EqR). The ARC-AGI-1 score is the pass@K logged during training, a vote pooled over the run's evaluations. The GitHub README describes the evaluation protocol and its caveats.

Download

From the root of a clone of the code repository:

# all checkpoints (4.0 GB)
hf download arberzela/sfprm --local-dir checkpoints

# evaluation weights only, without the resume bundles (2.6 GB)
hf download arberzela/sfprm --local-dir checkpoints --exclude "*_train_state.pt"

# a single checkpoint
hf download arberzela/sfprm --local-dir checkpoints --include "sfprm-maze-step78120/*"

Evaluation

After building the datasets as the code repository's README describes, evaluate.py scores a checkpoint directory with the training loop's own evaluation:

python evaluate.py checkpoints/sfprm-maze-step78120 --batch 96
python evaluate.py checkpoints/sfprm-miniarc-step68620 arch.max_iter_eval=200

Citation

@misc{zela2026sfprm,
      title={Spherical Fixed-Point Reasoners: Adaptive Loop Halting on the Hypersphere},
      author={Arber Zela and Artavazd Maranjyan and Pol Puigdemont and David Berghaus and Valentyn Boreiko and Volkan Cevher},
      year={2026},
}