Reproduction bundle — "Provably Data-driven Lagrangian Relaxation for Mixed Integer Linear Programming"

arXiv 2605.19052 · OpenReview OwLuqetJuB · reproduced by SabaPivot for the ICML 2026 Reproducibility Challenge. All six claims verified. CPU only, ~52 min, $0.

Logbook: https://huggingface.co/spaces/SabaPivot/repro-provably-data-driven-lagrangian-relaxation-for-mixed-integer-linear-programming

Layout

• scripts/common.py — flat random MILP family, exact dual oracle (enumeration + HiGHS cross-check), Kelley cutting planes
• scripts/blockfam.py — block-decomposable MILP family (instance grows with s)
• scripts/hardfam.py — the paper's own hard family + exact Bayes minimax lower bounds
• scripts/claim{1,1b,2,3,4,4b,5,6}*.py — one script per claim
• scripts/make_figures.py
• outputs/*.json — every raw number quoted in the logbook
• figs/*.png
• poster/ — posterly source, GATE_REPORT.json (PASS), rendered PDF/PNG, poster_embed.html

Rerun (CPU, no GPU, no API key)

pip install "numpy>=2" "scipy>=1.15" matplotlib
python scripts/claim6_prop51.py      #   4 s
python scripts/claim2_thm56.py       #  15 s
python scripts/claim1_thm55.py       #  26 min  (writes outputs/general_sweep.json)
python scripts/claim1b_blocks.py     # 9.7 min
python scripts/claim3_thm512.py      # 2.4 min  (reads outputs/general_sweep.json)
python scripts/claim4b_blocks.py     #  37 s
python scripts/claim4_thm61_62.py    # 4.3 min
python scripts/claim5_covering.py    # 8.7 min
python scripts/make_figures.py