DiscoGen: Procedural Generation of Algorithm Discovery Tasks in Machine Learning
This repository contains code for the DiscoGen modular benchmark for automated algorithm discovery.
- Paper: https://arxiv.org/pdf/2603.17863
- GitHub repository: https://github.com/AlexGoldie/discogen/
- Documentation: https://AlexGoldie.github.io/discogen/
Quick Start
Install DiscoGen:
pip install discogen
List available domains:
discogen get-domains
Create a task (the following will include a full implementation, and no editable modules):
discogen create-task --task-domain OnPolicyRL
Create an example task for running an agent (with an incomplete set of modules):
discogen create-task --task-domain OnPolicyRL --example
See the full documentation for detailed usage. Please note that each task_domain has its own set of requirements which may need to be installed, and these are often in conflict with the base DiscoGen requirements. This can be done using the install.sh script provided in each task folder.
Every domain includes references in discogen/domains/<task_domain>/utils/_reference.txt.
Task Domains
| Task Domain | Modules | Datasets | Description |
|---|---|---|---|
| BayesianOptimisation | acq_fn, acq_optimizer, sampler, next_queries, surrogate, surrogate_optimizer | Ackley1D, Ackley2D, Branin2d, Bukin2d, Cosine8d, DropWave2d, EggHolder2d, Griewank5d, Hartmann6d, HolderTable2d, Levy6d. | Optimization of black-box functions using surrogate models to find global minima/maxima. |
| BrainSpeechDetection | loss, networks, optim | 7 LibriBrainSherlock tasks. | Detecting speech features directly from brain activity data. |
| ComputerVisionClassification | loss, networks, optim, preprocess | CIFAR10, CIFAR10C, CIFAR10LT, CIFAR100, FashionMNIST, MNIST, OxfordFlowers, StanfordCars, TinyImageNet. | Image classification on a range of datasets. |
| ContinualLearning | optim, regularizer, replay, sampler, scheduler | PermutedMNIST, SplitCIFAR100, TinyImageNetSplit. | Training a model on continually changing data, such that it can adapt to new data without losing old capabilities. |
| GreenhouseGasPrediction | data_processing, model | 4 Mauna Loa Time-series (CO2, N2O, SF6, CH4). | Time-series forecasting of atmospheric greenhouse gas concentrations. |
| LanguageModelling | loss, networks, optim | OPCFineWebCode, OPCFineWebMath, LMFineWeb, TinyStories. | Training transformer-based models on code, mathematics, and narrative text. |
| ModelUnlearning | loss | MUSE, TOFU, WMDP_Cyber. | Fine-tuning pretrained models to remove specific knowledge or data points while retaining others. |
| NeuralCellularAutomata | loss, optimiser, perceive, train, update | GrowingButterfly, GrowingLizard, MatrixOperations, MNISTInpainting, SelfClassifyingMNIST. | Evolving neural cellular automata to different tasks based on reproduction and classification. |
| OfflineRL | actor_loss, critic_loss, optim, networks, train | 10 OGBench | Training RL policies from offline datasets. |
| OffPolicyRL | q_update, policy, networks, optim, rb, train, config | 4 MinAtar. | Value-based RL for training an agent in MinAtar. |
| OnPolicyMARL | activation, loss, networks, optim, targets, train | 5 MABrax, MPE Spread, 11 SMAX | Training multiple on-policy RL agents in different multi-agent environments. |
| OnPolicyRL | loss, networks, optim, train, activation, targets, schedule | 4 MinAtar, 7 Brax, 2 Craftax. | Training an RL agent in a range of different RL environments using PPO-style algorithms. |
| TrajectoryPrediction | loss, networks, optim, train | Argoverse2, nuScenes, Waymo. | Training a model to predict trajectories in autonomous driving datasets. |
| UnsupervisedEnvironmentDesign | sample_levels, train_step, variable_config | 3 Kinetix sizes, Minigrid. | Generating and curating training environments/levels to improve RL agent generalization. |
Development Setup
1. Set Up Your Development Environment
Install the environment and the pre-commit hooks with:
make install
This will also generate your uv.lock file.
Contributing
We welcome contributions! DiscoGen grows stronger with more tasks and domains.
- Found a bug? Open an issue
- Want to add a task? See our Contributing Guide
- Adding datasets? Check the Dataset Integration Guide
See CONTRIBUTING.md for detailed development guidelines.
Citation
If you use DiscoGen in your research, please cite:
@misc{goldie2026proceduralgenerationalgorithmdiscovery,
title={Procedural Generation of Algorithm Discovery Tasks in Machine Learning},
author={Alexander D. Goldie and Zilin Wang and Adrian Hayler and Deepak Nathani and Edan Toledo and Ken Thampiratwong and Aleksandra Kalisz and Michael Beukman and Alistair Letcher and Shashank Reddy and Clarisse Wibault and Theo Wolf and Charles O'Neill and Uljad Berdica and Nicholas Roberts and Saeed Rahmani and Hannah Erlebach and Roberta Raileanu and Shimon Whiteson and Jakob N. Foerster},
year={2026},
eprint={2603.17863},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2603.17863},
}
License
DiscoGen is released under the MIT License.
Metadata
Release files for discogen 1.3.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| discogen-1.3.0.tar.gz | 1.0 MB | Without provenance | Details for discogen-1.3.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| discogen-1.3.0-py3-none-any.whl | Python 3 | none | any | Details for discogen-1.3.0-py3-none-any.whl |
Total release size: 1.9 MB
Release files / discogen-1.3.0.tar.gz
| Download URL | discogen-1.3.0.tar.gz |
|---|---|
| Size | 1.0 MB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
e1237067c2a7938007c149a1aba9a996a8376eadcfdb210a0c44eb1f002bedb8
|
|
BLAKE2b-256 checksum How to use checksums |
00f8bcac0770e5996cb84acf4ac27a14211511febfadb1ac31ab0ce271d0702a
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
uv/0.6.14
|
Release files / discogen-1.3.0-py3-none-any.whl
| Download URL | discogen-1.3.0-py3-none-any.whl |
|---|---|
| Size | 855.6 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
2f3c24c07b5811cf24ecbdf6b609bb901cad349f709a6b90f0e621907dac2c46
|
|
BLAKE2b-256 checksum How to use checksums |
0d60fc8b435949df965dfbc4935fb36405a760869e213ca3f5b80fdf2b56d56b
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
uv/0.6.14
|
Security for discogen 1.3.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.