Reproduce “Transformers are Minimax Optimal Nonparametric In-Context Learners” 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

Juno Kim, Tai Nakamaki, Taiji Suzuki

The authors analyze in-context regression with a pretrained DNN feature map and single linear-attention layer, which uses prompt examples without weight updates. For tasks in a Besov function class, their empirical risk minimizer achieves query-prediction mean-squared error at the minimax rate, up to a logarithmic factor, given sufficient pretraining tasks. This theoretical result applies under stated assumptions, not to language models.

Summary by alphaXiv.

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