Reproduce “When Teachers Mislead: Spurious-Signal-Aware 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
Yinuo Jiang, Yongjie Ye, Zhou Tao, Xiang Zhuang, Qiang Zhang, Huajun Chen, Tiankai Li
Jiang et al. (Zhejiang University, ByteDance) developed Spurious-Signal-Aware On-Policy Distillation (SA-OPD), a framework that identifies and filters misleading 'spurious signals' from teacher models during knowledge transfer. SA-OPD improved average scores by 3.5 points in VLM visual understanding and 1.9 points in LLM mathematical reasoning tasks compared to baseline methods, by focusing on input-grounded teacher supervision.
Key insights
- Teacher models can generate 'spurious signals' that provide optimization-relevant supervision not grounded in the specific task input, hindering effective knowledge transfer to student models.
- Assessing the input-groundedness of teacher signals, in addition to metrics like confidence or divergence, is crucial for selecting high-quality supervision in On-Policy Distillation.
- Filtering out tokens that show both weak input-groundedness and strong optimization impact can stabilize training and lead to more robust student models less susceptible to teacher biases.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.