Reproduce “DISCO Balances the Scales: Adaptive Domain- and Difficulty-Aware Reinforcement Learning on Imbalanced Data” 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
Yuhang Zhou, Jing Zhu, Shengyi Qian, Zhuokai Zhao, Xiyao Wang, Xiaoyu Liu, Ming Li, Paiheng Xu and 2 more
DISCO modifies Group Relative Policy Optimization (GRPO) for skewed multi-domain training by scaling each prompt group’s rewards according to domain rarity and answer self-consistency, then centering the scaled rewards without standard-deviation normalization. With Qwen3-0.6B, it reaches 40.86% mean exact-match accuracy versus 36.46% for Naive GRPO across four heavy mixtures, though gains and trade-offs vary by model and task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.