Reproduce “Avoiding Obfuscation with Prover-Estimator Debate” 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
Jonah Brown-Cohen, Geoffrey Irving, Georgios Piliouras, Lijie Chen, Jiawei Li, Zhiyang Xun
The paper introduces prover-estimator debate, in which a proposer decomposes an answer into subclaims, an estimator assigns probabilities to them, and the proposer chooses one estimate to challenge. The authors prove that, under an A-provable, rho-stable decomposition and an accurate A-leading Stackelberg equilibrium, answer accuracy exceeds 1-5 epsilon. This is a theoretical, distributional guarantee, not benchmark or worst-case evidence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.