ludwigw/beluga-trm-105m

Beluga TRM (105.8M)

TRM for Airbus Beluga XL logistics constraint satisfaction solving (2,336 real-world problems).

Model Details

• Parameters: 105.8M
• Task: Multi-constraint optimization for aircraft loading
• Input: Variable-dimension (69-821 jigs, 43-199 flights)
• Constraints: Flight capacity, rack capacity, scheduling, type matching

Architecture

TinyRecursiveMLP(
    x_dim=dynamic,  # Padded to max dimensions
    y_dim=512,
    z_dim=512,
    hidden=1024,
    num_classes=max_jigs * max_flights,  # Assignment matrix
    H_cycles=2,
    L_cycles=2
)

Performance

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

Real-World Application

Aerospace logistics planning with certified constraint satisfaction. Handles:

• Dynamic padding/masking for variable dimensions
• Multi-objective optimization (4 constraint types)
• Safety-critical deployment requirements

Usage

import torch
from veriphi.models import TinyRecursiveMLP
from veriphi.data import BelugaDataset

# Load model
model = TinyRecursiveMLP(...)  # See architecture above
model.load_state_dict(torch.load("beluga-trm-105m.pt"))
model.eval()

# Load problem
dataset = BelugaDataset("data/beluga/deterministic")
state_tensor, problem = dataset[0]

# Solve
with torch.no_grad():
    assignment_logits = model(state_tensor)
    assignment = assignment_logits.reshape(problem.num_jigs, problem.num_flights)

Dataset

TUPLES Beluga AI Challenge dataset (2,336 problems):

• Training: 1,869 problems
• Validation: 467 problems
• Complex multi-constraint optimization

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