Reproduce “AdaC-GRPO: Towards Mastering Long-Horizon Visual Sudoku Puzzles via Adaptive Curriculum Learning” 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
Sida Li, Yuanxin Liu, Linli Yao, Shicheng Li, Xu Sun, Haoli Bai, Xiaohui Li
AdaC-GRPO selects Sudoku difficulty groups whose measured exact-answer accuracy falls between 0.30 and 0.80, then trains with group-relative policy updates. On FG-Sudoku, Qwen2.5-VL-7B reaches Pass@1 0.586 overall and 0.219 on the aggregate 9×9 group, compared with DAPO’s 0.540 and 0.086. This is a benchmark result, not parity with proprietary models or a guarantee of informative feedback on each prompt.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.