Reproduce “Beyond Uncertainty: Multi-Solver Disagreement Rewards for Self-Evolving Reasoning Curricula” 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
Vinoth Selvendran, Zhanming Zhang
The method adds cross-variant disagreement to the uncertainty reward for training a Challenger in a self-evolving reasoning system. It compares plurality answers from three solver variants while retaining pooled-answer uncertainty. Its Qwen3-4B Solver reaches a 45.69% average on MATH-500, AMC, and Olympiad, versus 44.35% for R-Zero’s iteration-1 Solver, but scores lower on GSM8K; this comparison does not isolate disagreement’s effect.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.