Reproduce “Unlocking the Unsolvable: Teacher-Guided Curriculum for Data-Efficient RLVR” 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
Yukang Zhu, Zhen Han
A teacher-guided curriculum turns initially unsolvable mathematics problems into usable RLVR training examples by gradually removing partial reasoning hints. The proposed Monotone Frontier Curriculum (MFC) improves average benchmark performance over standard GRPO while using 128 source problems instead of 2,000, and expands high-sampling solution coverage by about 10 percentage points on AIME problems for both tested model sizes.
Key insights
- Teacher traces act as temporary scaffolding rather than supervised targets: varying their length creates a continuous difficulty landscape and restores reward variance on otherwise all-failed examples.
- A success-only, monotone reduction in guidance keeps moving problems toward their original form, unlike curricula that settle at a fixed band of hint difficulty.
- The gains are strongest at larger sampling budgets, suggesting that MFC broadens the model’s set of discoverable solution paths rather than merely concentrating probability on familiar successful answers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.