MaximeILL/GAR_ckpts

GAR checkpoints

The two Tiny Recursive Model (TRM) checkpoints on which the paper Resample the Answer: Where to Spend Test-Time Compute in Recursive Reasoning (Illouz, 2026) evaluates Gated Answer Resampling (GAR) and all the experiments.

File	Model	Benchmark	Checkpoint	Accuracy (standard TRM inference)
TRM-MLP-Sudoku-Extreme.pt	TRM-MLP, 5.03M parameters	Sudoku-Extreme	step 65,100, the last of 10 (50,000 epochs)	85.90% on the paper's 1,000 test puzzles, 86.50% on the full test split
TRM-Att-PPBench.pt	TRM-Att, 7.27M parameters	PPBench, six puzzle types	step 254,961, the last of 21 (21,000 epochs)	68.83% on validation (308 puzzles of the five reported types), 67.35% on golden (49 puzzles)

Both were trained with the released TRM code (Jolicoeur-Martineau, 2025), with AdamW instead of AdamATan2: TRM-MLP with the released command on 1,000 Sudoku-Extreme puzzles × 1,001 examples, and TRM-Att with the Sudoku-Extreme settings on 9,325 PPBench puzzles × 11 examples.

Files

Each .pt file is the PyTorch state dict of the network (EMA weights), saved by the TRM training code as step_65100 and step_254961 respectively. The keys start with _orig_mod., as the training code compiles the model with torch.compile. Each .yaml file is the configuration of its training run (the all_config.yaml written by the TRM training code).

Loading

import torch
from huggingface_hub import hf_hub_download

path = hf_hub_download("MaximeILL/GAR_ckpts", "TRM-MLP-Sudoku-Extreme.pt")
state_dict = torch.load(path, map_location="cpu")

Citation

@misc{illouz2026resample,
  title         = {Resample the Answer: Where to Spend Test-Time Compute in Recursive Reasoning},
  author        = {Maxime Illouz},
  year          = {2026},
  eprint        = {2610.16427},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2610.16427}
}