Reproduce “RPSD: Retrieval-Privileged On-Policy Self-Distillation with Optimal Transport” 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
Dongyue Li, Zebang Cheng, Zhongyi Deng, Hongboxu, Yi Bin, Fei Ma
RPSD trains next-item recommenders by using similar users’ interaction sequences as training-only teacher evidence. It re-scores sampled answers with ordinary and evidence-enriched contexts, then uses token-level confidence differences to redistribute reward-based credit. On Netflix, using Qwen3-4B and ten sampled candidates, it achieves HR@5 of 0.9245 and NDCG@5 of 0.7448, above GRPO, but this sampled-candidate result does not establish full-catalog performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.