Reproduce “Steerable but Not Decodable: Function Vectors Operate Beyond the Logit Lens” 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
Mohammed Suhail B Nadaf
The authors test whether function-vector steering makes answers readable through a model’s vocabulary readout. They compare generated-task accuracy after adding a task-derived vector to intermediate states with logit-lens top-10 accuracy. Across 72 task-by-model cells, 10 are steerable but not logit-lens-decodable, while 3 show the reverse; a nonlinear probe recovers input-conditional information in 5 of those 10 cases.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.