Reproduce “On the Primacy Bias in RLVR Training” 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

Phuc Minh Nguyen, Ngoc-Hieu Nguyen, Parshin Shojaee, Chinh Duc La, Duy Minh Ho Nguyen, Rui Zhang, Khoa D. Doan, Ting Hua and 1 more

The authors study whether RLVR expands math reasoning coverage or sharpens existing abilities. They train Qwen2.5-3B-Instruct with GRPO on difficulty-selected MATH examples and compare Pass@128 overall and on problems with zero observed base-model accuracy. Hard-only training scores 69.87 on OlympiadBench, above the base’s 68.77, but scores 52.98 on Minerva, below the base’s 53.91. Gains are not universal.

Summary by alphaXiv.

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