Reproduce “Slow-Fast Multi-Teacher On-Policy Distillation for Capability Preservation” 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
Xiaofei Yin, Tong Chu, Jiyuan Fu, Jun Lan, Shuheng Zhou, Huijia Zhu
SF-MOPD lets a multimodal student learn from domain specialists while filtering teacher-suggested changes that widen its distance from a slowly updated exponential moving average reference. On Qwen3-VL-8B, it raises average accuracy across five general benchmarks from 73.3% to 76.4% and six domain-specific benchmarks from 59.0% to 60.4%, compared with vanilla MOPD. The general average remains below the instruct model’s 77.2%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.