Reproduce “On the Emergence of Cross-Task Linearity in the Pretraining-Finetuning Paradigm” 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
Zhanpeng Zhou, Zijun Chen, Yilan Chen, Bo Zhang, Junchi Yan
This work introduces Cross-Task Linearity (CTL), a phenomenon where the features of linearly interpolated models, finetuned from a common pretrained checkpoint on different tasks, are approximately equal to the linear interpolation of their individual features. This property offers feature-centric explanations for the effectiveness of model merging and task arithmetic.
Key insights
- Models finetuned from a shared pretrained checkpoint on different tasks exhibit an approximate linear relationship in their feature space (CTL), even when parameter-space linear connectivity (LMC) does not hold.
- The efficacy of model averaging can be directly understood as averaging features across models, while task arithmetic operations (addition, negation) correspond to adding or subtracting feature information.
- The emergence and strength of CTL are significantly influenced by the extent of pretraining and the similarity between pretraining and finetuning tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.