This paper critiques GRPO's bias in training language models for theorem proving and introduces the unlikeliness reward to enhance performance and sample diversity, achieving competitive results.

https://arxiv.org/abs//2506.02355

YouTube: https://www.youtube.com/@ArxivPapers

TikTok: https://www.tiktok.com/@arxiv_papers

Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476016

Spotify: https://podcasters.spotify.com/pod/show/arxiv-papers