Reproduce “Language Model Activations Inhabit Privileged Error-Correcting Basins” 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
Matthew Finlayson, Francisco Pernice, Eric Todd, Amir Zur, Daniel Wurgaft, Fenil R. Doshi, Vasudev Shyam, Matt Feiszli and 11 more
The authors test whether layer geometry can guide activation steering in an English-to-Chinese task. They identify low-expansion activation basins and choose a per-instance steering strength from a velocity minimum, with a tuned-strength fallback. On held-out FineWeb data using Llama 3.1, adaptive steering raises Chinese greedy outputs from 74.81% to 77.94%, in this tested setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.