Reproduce “Learning to Program Retrievals: Composing Search Primitives via Reinforcement Learning” 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
Kyuyoung Kim, Zilin Xiao, Zhiyuan Wang, Avinash Atreya, Ke Yang, Jinwoo Shin, Kecheng Hao
RPL lets question-answering agents write executable search programs that retrieve and process candidates before selecting evidence for the model’s context. To address sparse reinforcement-learning feedback, the authors mutate and verify program variations, then use them for distillation. On seven held-out open-domain QA benchmarks, Qwen3-4B averages 0.515 exact-match accuracy versus Search-R1’s 0.469; this average does not imply gains on every task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.