Neural Low-Discrepancy Sequences

Paper • ICML 2026

Python 3.10+ License: MIT

This repository contains the code for our ICML 2026 paper Neural Low-Discrepancy Sequences. It provides PyTorch implementations for model training, discrepancy-based fine‐tuning, and reproducible experiments.

NeuroLDS architecture

Figure: Overview of the NeuroLDS architecture.

• scripts/models.py — Main model definition (NeuroLDS).
• scripts/main.py — Create/generate sequences; trains/evaluates NeuroLDS.
• scripts/utils.py — Aid/utility functions (discrepancy losses, seeding, I/O, plotting).
• scripts/hypertuning.py — Hyperparameter optimization via Optuna.
• scripts/smoke_test.sh — Sanity check; run to verify the setup works end-to-end.

# Recommended: Python 3.10–3.12 (tested on 3.11)
python3.11 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

Run the smoke test to verify that everything is set up correctly:

cd scripts
bash smoke_test.sh

This project is licensed under the MIT License - see the LICENSE file for details.