ludwigw/trm-mnist-baseline

TRM MNIST Baseline

Tiny Recursive Model (105.8M parameters) trained on MNIST without robustness methods.

Model Details

• Architecture: TRM-MLP (x_dim=784, y_dim=512, z_dim=512, hidden=1024, H=2, L=2)
• Parameters: 105.8M
• Training: Standard cross-entropy, 20 epochs, lr=1e-3
• Dataset: MNIST (60k train, 10k test)

Verification Results (β-CROWN, 512 samples)

ε (L∞)	Verified	Time/sample
0.01	12%	0.15s
0.03	3%	0.18s
0.06	1%	0.21s

Key Finding: Baseline fails beyond ε=0.01.

Usage

import torch
from veriphi.models import TinyRecursiveMLP

model = TinyRecursiveMLP(x_dim=784, y_dim=512, z_dim=512, hidden=1024, 
                         num_classes=10, H_cycles=2, L_cycles=2)
model.load_state_dict(torch.load("trm-mnist-baseline.pt"))
model.eval()

x = torch.randn(1, 784)
logits = model(x)

Citation

@article{deshmukh2026veriphi,
  title={Veriphi: Attack-Guided Neural Network Verification with Dataset-Dependent Training Methods},
  author={Deshmukh, Pratik and Savin, Vasili and Arya, Kartik},
  journal={arXiv preprint arXiv:2606.18454},
  year={2026}
}

Paper: arXiv:2606.18454 | Code: GitHub