Views
No views yet
1TinyRecursiveMLP(
2 x_dim=dynamic, # Padded to max dimensions
3 y_dim=512,
4 z_dim=512,
5 hidden=1024,
6 num_classes=max_jigs * max_flights, # Assignment matrix
7 H_cycles=2,
8 L_cycles=2
9)| Metric | Value |
|---|---|
| Training Loss | 930 → 2.26 (99.8% reduction) |
| Constraint Violations | Near-zero on validation |
| Inference Time | 2.6s per problem |
| Verification Time | 5× faster with attack-guided approach |
1import torch
2from veriphi.models import TinyRecursiveMLP
3from veriphi.data import BelugaDataset
4
5# Load model
6model = TinyRecursiveMLP(...) # See architecture above
7model.load_state_dict(torch.load("beluga-trm-105m.pt"))
8model.eval()
9
10# Load problem
11dataset = BelugaDataset("data/beluga/deterministic")
12state_tensor, problem = dataset[0]
13
14# Solve
15with torch.no_grad():
16 assignment_logits = model(state_tensor)
17 assignment = assignment_logits.reshape(problem.num_jigs, problem.num_flights)1@article{deshmukh2026veriphi,
2 title={Veriphi: Attack-Guided Neural Network Verification with Dataset-Dependent Training Methods},
3 author={Deshmukh, Pratik and Savin, Vasili and Arya, Kartik},
4 journal={arXiv preprint arXiv:2606.18454},
5 year={2026}
6}