Published Pareto-front Models

The final, completely verified per-method Pareto fronts from the ICML 2026 paper “Rethinking Evaluation Paradigms in IBP-based Certified Training” are available in the kkaulen/ctrain_pareto_fronts Hugging Face model repository.

The repository manifest records each model's dataset, architecture, training method, perturbation radius, reported complete-verification accuracies, and SHA-256 checksum. CTRAIN's paper utility uses this metadata to select the correct architecture and wrapper before loading a checkpoint.

MTL-IBP CIFAR-10 Example

The runnable example script prints the canonical MTL-IBP CNN7 front at epsilon 2/255, selects its highest-certified member, downloads it, and evaluates it on the CIFAR-10 test set:

python papers/rethinking_evaluation_paradigms/examples/evaluate_mtl_front.py

By default, the script evaluates all 10,000 test examples with IBP certification and CTRAIN's PGD attack. CUDA is strongly recommended. Common variants are:

# Inspect the manifest front without downloading weights.
python papers/rethinking_evaluation_paradigms/examples/evaluate_mtl_front.py \
  --list-only

# Evaluate a shorter test-set prefix.
python papers/rethinking_evaluation_paradigms/examples/evaluate_mtl_front.py \
  --test-samples 100

# Select another displayed front member.
python papers/rethinking_evaluation_paradigms/examples/evaluate_mtl_front.py \
  --index 2

The script prints the paper's complete-verification metrics separately from the fresh CTRAIN evaluation. The latter uses incomplete certification through CTRAINWrapper.evaluate; selecting ADAPTIVE combines IBP, CROWN-IBP, and CROWN rather than reproducing the paper's complete alpha-beta-CROWN run.

For publication details and full reproduction instructions, see the paper artifacts.