Reproduce “The Assistance Dilemma: Learning to Teach via Multi-Turn Reinforcement 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
Jakub Macina, Manu Kapur, Mrinmaya Sachan
Eduardo trains an AI math tutor using a reward based on whether a student solves a changed problem after tutoring, with tutor turns hidden. Separate gates penalize incorrect help and solution handover. In evaluations, Eduardo-27B's MathTutorBench aggregate score was 0.79, near Gemini 3.1 Pro at 0.80. These results measure benchmarked model behavior, not learning gains among human students.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.