Reproduce “Hierarchical Credit Assignment for RLVR on Fused Gromov-Wasserstein Geometry” 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

Qi Yu, Ruizhong Qiu, Zhichen Zeng, Xuying Ning, Yanjun Zhao, Dongqi Fu, Yinglong Xia, Hong Li and 1 more

HARA reshapes token-level credit in group-based reinforcement learning with verifiable rewards by comparing each rollout’s hidden states and token positions with a shared reference for rollouts with the same outcome. It uses fused Gromov–Wasserstein barycenters to derive novelty weights. On Qwen3-8B-Base, HARA-tok scores 63.9 average versus 59.1 for GRPO across five math benchmarks; the result is limited to math evaluations.

Summary by alphaXiv.

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