BIGR: frozen TRM backbones
Tiny recursive models, think BIGR.
Assets for the NeurIPS 2026 paper Boosting Inference with Guided Reasoning: Stochastic Exploration for Recursive Models (Corbett, Tzatzopoulou, Sood, Ramesh, Dodwell). Code and reproduction recipes: https://github.com/digiLab-ai/bigr. Paper: https://openreview.net/forum?id=nZD0FHMUmR.
This repository mirrors the data/ tree that the code expects, so
python scripts/download_checkpoints.py places every file where the experiment scripts look:
sudoku_extreme/checkpoints/last.ckpt Sudoku-Extreme backbone (TRM, 7M parameters, 74 MB)
sudoku_extreme/checkpoints/epoch_*.ckpt the 20 intermediate checkpoints of the same run (Figure 8)
maze_hard/checkpoints/last.ckpt Maze-Hard backbone (104 MB)
ppbench/checkpoints/last.ckpt PPBench backbone (106 MB)
The checkpoints are PyTorch Lightning checkpoints of the vendored nano-trm TRMModule
(EMA weights included); the code loads them with bigr.load_trm_backbone_checkpoint.
Nothing is trained for inference: the Q-head that BIGR uses as its guide is part of every
checkpoint. Training configs and commands are in the code repository.
The evaluation splits the paper scores (Sudoku-Extreme Test 10,020 and DF 1,427; Maze-Hard
Test 1,000 and DF 134) are not redistributed: python scripts/data/build_eval_cohorts.py in
the code repository rebuilds them from the Hierarchical Reasoning Model releases
sapientinc/sudoku-extreme and sapientinc/maze-30x30-hard-1k.
The checkpoints are released under CC BY 4.0: credit digiLab Solutions Ltd and cite the
paper when you use them. NOTICES.md in the code repository lists the licences of the
datasets and of the backbone code they were trained with.