Reproduce “Uncovering the Turing Test's Latent Signal: Indistinguishability as a Reward” 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
Emanuele Rimoldi, Riccardo Neumarker, Tomaso Poggio, Daniel Mitropolsky
The authors train a small model on four checkable tasks using a stronger model’s probability that it could have produced each student answer as a reward. They use GRPO without teacher-generated answers or correctness labels in updates. Held-out accuracy rises in all four tasks, including GSM8K from 55.7% to 68.8%; evidence is limited to one student, four datasets, and two teachers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.