Reproduce “Lexicographic Multi-Objective On-Policy Distillation” 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
Doseok Jang, Jon Ander Campos, Youran Qi
LMOPD combines reward-specialized math assistants by routing each student-generated answer to the specialist for its highest-priority unmet requirement. It adjusts token-level signals to reduce conflicts with earlier priorities. In a four-expert test on a roughly 30-billion-parameter mixture-of-experts model, it retained 90.6% of specialist accuracy gain and 88.6% of reasoning-correctness gain, exceeding unified baselines, but verification and efficiency gains were negative; evidence is limited to math reasoning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.