Reproduce “OR-Transformer: Scaling Real-Time Decision-Making to 1,000 Items” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Shuze Daniel Liu, David Simchi-Levi, Claire Chen, Chutong Gao, Shangtong Zhang

OR-Transformer addresses joint replenishment by representing products as an unordered set and learning shared order-opening decisions separately from item-level quantities. Quantities receive gradients through inventory transitions and future costs. In synthetic tests up to 1,024 items, it reports cost 0.35 million versus PPO’s 1.39 million. This benchmark does not establish field performance or isolate which component drives gains.

Key insights

  • Separating discrete and continuous gradient estimation provides more direct learning signals for replenishment quantities than applying policy gradients uniformly to the full action space.
  • Self-attention lets each item condition its replenishment decision on other products, capturing shared setup costs and demand correlations while preserving invariance to item ordering.
  • The main advantage emerges at scale: learning methods that remain competitive for small catalogs can diverge or degrade sharply as the number of items increases.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.