Reproduce “The Geometry of Inference in Transformer Residual Streams” 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
Timur Mudarisov, Mikhail Burtsev, Radu State
The authors measure how transformer states compare with their own final residual representations and endpoints from other contexts. They find own endpoints are closer on average from the earliest measured post-block state, although individual alternatives often remain closer. A held-out endpoint-distribution test predicts cosine competitor fractions better than a uniform-sphere baseline, but this geometry is bank- and metric-dependent.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.