Reproduce “Leverage Is Not Reach: A Control-Window Law for Single-Neuron Steering in Language Models” 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
Hongliang Liu
The authors define a budgeted control window for neuron interventions: a dose must cross a behavior’s trigger before reaching the ceiling where coherent output collapses. They predict the ceiling from weights and a generic forward pass, but measure triggers through rollouts. Across 15 held-out safety controllers, ceiling error averaged 0.14; 11/15 verdicts were correct, barely above the 10/15 baseline.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.