URM - GitLite

KO | EN

gitlite — search

#python #llm #go #rust #deep-learning #linux #docker #nodejs #windows #awesome #cli #kubernetes #awesome-list #openai

URM

(★ 134)

Universal Reasoning Model

File Explorer

Download Latest Version (.zip)

- - hrm.yaml
  - trm.yaml
  - urm.yaml
- cfg_eval.yaml
- cfg_pretrain.yaml
- build_arc_dataset.py
- build_maze_dataset.py
- build_sudoku_dataset.py
- common.py
- - arc.cpython-311.pyc
- arc.py
- arc_augmentation_ablation.py
- - arc-agi_concept-challenges.json
  - arc-agi_concept-solutions.json
  - arc-agi_evaluation-challenges.json
  - arc-agi_evaluation-solutions.json
  - arc-agi_evaluation2-challenges.json
  - arc-agi_evaluation2-solutions.json
  - arc-agi_training-challenges.json
  - arc-agi_training-solutions.json
  - arc-agi_training2-challenges.json
  - arc-agi_training2-solutions.json
- - hrm_act_v1.py
  - hrm_act_v2.py
- - trm.py
- - urm.py
- common.py
- layers.py
- losses.py
- muon.py
- sparse_embedding.py
- URM_arcagi1.sh
- URM_arcagi2.sh
- URM_sudoku.sh
attn_maps_ab.py
evaluate_trained_model.py
pretrain.py
puzzle_dataset.py
README.md
requirements.txt
utils.py

# Installation Guide

1. Get the code

git clone https://github.com/UbiquantAI/URM

Downloads the entire project code from GitHub to your computer.

cd URM

Moves into the project folder you just downloaded.

2. Python

Easy Recommended

Prerequisites

Git Needed to download the project code from GitHub.
Python 3 On Windows, be sure to check 'Add Python to PATH' during installation.

pip install -r requirements.txt

Installs the Python libraries listed in requirements.txt (or similar).

python -m data.build_arc_dataset \

Runs the Python script (or module).

✅ If it runs without errors and prints output in the terminal, it worked.

Pulled directly from this repo's README.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Universal Reasoning Model

[![paper](https://img.shields.io/badge/paper-A42C25?style=for-the-badge&logo=arxiv&logoColor=white)](https://arxiv.org/abs/2512.14693)

Universal transformers (UTs) have been widely used for complex reasoning tasks such as ARC-AGI and Sudoku, yet the specific sources of their performance gains remain underexplored. In this work, we systematically analyze UTs variants and show that improvements on ARC-AGI primarily arise from the recurrent inductive bias and strong nonlinear components of Transformer, rather than from elaborate architectural designs. Motivated by this finding, we propose the Universal Reasoning Model (URM), which enhances the UT with short convolution and truncated backpropagation. Our approach substantially improves reasoning performance, achieving state-of-the-art 53.8% pass@1 on ARC-AGI 1 and 16.0% pass@1 on ARC-AGI 2.

## ⚠️ For Question Regarding Sudoku Score
The reported score of 87.4% in the TRM paper is obtained using an MLP model, which we believe it is completely different from the TRM architecture in ARC-AGI task. Therefore, **_for fair comparison_**, when reproducing the results, we unified the architectures for ARC-AGI 1, ARC-AGI 2, and Sudoku to be exactly the same, which means **the architecture used to reproduce Sudoku is the same TRM architecture used to run ARC-AGI**.

Reproducing the correct TRM Sudoku score:
```bash
git clone https://github.com/SamsungSAILMontreal/TinyRecursiveModels
cd TinyRecursiveModels
python dataset/build_sudoku_dataset.py --output-dir data/sudoku-extreme-1k-aug-1000  --subsample-size 1000 --num-aug 1000

run_name="pretrain_att_sudoku"
python pretrain.py \
arch=trm \
data_paths="[data/sudoku-extreme-1k-aug-1000]" \
evaluators="[]" \
epochs=50000 eval_interval=5000 \
lr=1e-4 puzzle_emb_lr=1e-4 weight_decay=1.0 puzzle_emb_weight_decay=1.0 \
arch.L_layers=2 \
arch.H_cycles=3 arch.L_cycles=6 \
+run_name=${run_name} ema=True
```

Results:

<img width="400" height="250" alt="image" src="https://github.com/user-attachments/assets/c0699d98-64d1-4f41-9c8f-818f924ad77b" />

## Installation
```bash
pip install -r requirements.txt
```

## Login Wandb
```bash
wandb login YOUR_API_KEY
```

## Preparing Data
```bash
# ARC-AGI-1
python -m data.build_arc_dataset \
  --input-file-prefix kaggle/combined/arc-agi \
  --output-dir data/arc1concept-aug-1000 \
  --subsets training evaluation concept \
  --test-set-name evaluation

# ARC-AGI-2
python -m data.build_arc_dataset \
  --input-file-prefix kaggle/combined/arc-agi \
  --output-dir data/arc2concept-aug-1000 \
  --subsets training2 evaluation2 concept \
  --test-set-name evaluation2

# Sudoku
python data/build_sudoku_dataset.py --output-dir data/sudoku-extreme-1k-aug-1000  --subsample-size 1000 --num-aug 1000
```

## Reproducing ARC-AGI 1 Score
```bash
bash scripts/URM_arcagi1.sh
```

## Reproducing ARC-AGI 2 Score
```bash
bash scripts/URM_arcagi2.sh
```

## Reproducing Sudoku Score
```bash
bash scripts/URM_sudoku.sh
```

### Citation
```
@misc{gao2025universalreasoningmodel,
      title={Universal Reasoning Model}, 
      author={Zitian Gao and Lynx Chen and Yihao Xiao and He Xing and Ran Tao and Haoming Luo and Joey Zhou and Bryan Dai},
      year={2025},
      eprint={2512.14693},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2512.14693}, 
}
```