Reproduce “Language, Not Code Abstraction: Natural-language Training Shapes Brain-aligned Geometry” 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
Iñigo Parra
The authors compare GPT-2-small models trained on narrative English or four code corpora with matched random-weight models. On Pereira-243, mean held-out Pearson r across 12 blocks is 0.122 for TinyStories and 0.069 for the equally weighted code mean. English also outperforms code on Tuckute2024, but not every benchmark; corrected tests establish no TinyStories advantage over random weights.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.