Reproduce “Nash-MOPD: Optimizing Task Trade-offs in Multi-Teacher 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

Siyuan Gan, Bo Yang, Linjian Meng, Boyan Wang, Xiaolong Qi, Jing Huo, Yang Gao

Nash-MOPD adapts prompt allocation during multi-teacher on-policy distillation using smoothed reverse-KL gaps and estimated proportional progress per prompt. Its local frozen-rate model allocates more exposure to domains with slower estimated progress. In the DeepSeek-R1-Distill-Qwen-1.5B experiment, after 100 updates it exceeds both baselines’ 200-update pass@1 scores on MATH-500 and LiveCodeBench v6; this shows update, not measured wall-clock, savings.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.