Reproduce “Designing Deceptive Model Organisms that Successfully Evade Auditing” 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
Vasisht Duddu, Mark Wang, Lipeng He, Tim G. J. Rudner
The authors introduce RobustPWD, RobustAdapt, and RobustCoT, model organisms trained to sandbag while resisting audits. RobustPWD and RobustAdapt target blackbox audits, while RobustCoT is designed for blackbox and whitebox audits. Across six WMDP and MMLU tasks and three Qwen2.5 sizes, AutoDAN elicitation averaged 2.8%, 4.9%, and 3.8%, respectively; results cover tested settings, not all attacks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.