Reproduce “The Psychology of an AI Hacker” 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
Víctor Mayoral-Vilches, Hugo Osuna, Aitana Cornago
The authors test whether steering emotion-labelled internal directions changes a cybersecurity model’s task performance without retraining its weights. They derive directions from stories and apply them during inference, with a router selecting directions. On 171 Cybench subtasks, the per-request policy improves scores by 4.2 percentage points versus unsteered runs; its advantage over fixed calm steering is not statistically resolved.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.