Reproduce “GeoMin: Data-Efficient Semi-Supervised RLVR via Geometric Distribution Modeling” 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
Guangcheng Zhu, Shenzhi Yang, Haobo Wang, Xing Zheng, Yingfan MA, Xuening Feng, Zhongqi Chen, Kai Tang and 4 more
GeoMin is a semi-supervised Reinforcement Learning with Verifiable Rewards (RLVR) framework that models geometric distributions in LLM latent spaces to efficiently leverage unlabeled data. The method achieves superior performance, outperforming a fully supervised baseline with 10% of annotations, while demonstrating a 2.14x speedup over a competitive semi-supervised baseline.
Key insights
- A "geometric resonance" phenomenon reveals clear directional separation between correct and incorrect rollouts in LLM latent representations, which can be modeled.
- Modeling global feature distributions with vMF provides a robust prior for calibrating the veracity of self-reward signals on unlabeled data.
- Prioritizing and amplifying advantages for "ambiguous boundary samples" accelerates the establishment of a sharp geometric decision boundary early in training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.