Reproduce “Reward-Weighted Preference Optimization for Selective Combinatorial Optimization Problems” 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

Haohui Zhang, Xinyu Hu, Wouter van Heeswijk, Neil Yorke-Smith, Martijn Mes

RWPO trains neural solvers for selective combinatorial optimization by weighting preference comparisons according to reward-gap size, so larger improvements count more than near-ties. On out-of-distribution OP at 2,000 nodes, RWPO reports objective 134.607 and gap 19.28%, versus PO’s 49.028 and 70.60%; adapted RWPO+TS exceeds Compass’s mean reward from 100 to 2,000 nodes. It requires exact rewards.

Summary by alphaXiv.

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