REIGN

REIGN: Robust Expected Information Gain for Navigating Adaptive Perturbation Screens.

Paper | Reproducibility | Experiments | Citation metadata

REIGN is the camera-ready workshop repo for the paper REIGN: Robust Expected Information Gain for Navigating Adaptive Perturbation Screens. It contains the public package, the frozen benchmark artifacts used in the paper, shipped oracle caches, and the final PDF.

What Is Included

• src/reign/ - package code for surrogates, planners, and experiment runners
• paper/ - final workshop PDF
• paper/artifacts/ - frozen workshop tables and figures
• configs/ - workshop benchmark configs
• scripts/ - workshop reproduction and artifact scripts
• data/cache/ - shipped oracle caches used by the workshop experiments

Quickstart

uv sync --all-groups
uv run reign info
make check

The checks badge reflects this release target: ruff format --check, ruff check, and pytest all pass.

If you prefer a standard Python workflow, see CONTRIBUTING.md for a venv + pip install -e . setup.

Reproduce The Workshop Benchmarks

uv run python scripts/run_sweeps.py \
  --config configs/buchwald_hartwig_stream.yaml \
  --seeds 0:50 \
  --planners random greedy_mean bo_ei bo_ucb reign \
  --output-dir runs/buchwald

uv run python scripts/make_paper_artifacts.py \
  --runs runs/buchwald \
  --out outputs/buchwald_artifacts

See EXPERIMENTS.md and REPRODUCIBILITY.md for the full workflow.

What The Paper Shows

• on Buchwald-Hartwig, REIGN is competitive with the strongest BO baseline and stronger than greedy transfer, Thompson sampling, and MES
• when cross-context transfer is disabled, all methods collapse to random-level performance
• on low-context biological screens, simpler methods remain competitive
• the main practical gain comes from the hierarchical surrogate, and the acquisition helps most in the richer many-context regime

Data Credit

• The Buchwald cache is derived from rxn4chemistry/rxn_yields.
• The Buchwald benchmark source is Ahneman et al., Science 2018.
• The upstream rxn4chemistry/rxn_yields repository is MIT-licensed.

How To Cite

@misc{thomas2026reign_workshop,
  title = {REIGN: Robust Expected Information Gain for Navigating Adaptive Perturbation Screens},
  author = {Thomas, Noel},
  year = {2026},
  note = {MLGenX 2026 Workshop release},
  url = {https://github.com/11NOel11/REIGN}
}

@software{thomas2026reign,
  author = {Thomas, Noel},
  title = {REIGN},
  year = {2026},
  version = {1.0.0},
  url = {https://github.com/11NOel11/REIGN}
}

The machine-readable metadata is in CITATION.cff.

Release

• public release tag: REIGN-v1