Reproduce “Discovering Jailbreaks via Curriculum of 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
Michał Burzyński, Noah T. Rossi, Shomit Basu, Andy Arditi
CoM trains an investigator model to elicit refused behaviors by moving it through progressively less refusal-weakened versions of a target, advancing when mean reward reaches a threshold. Against the original gpt-oss-120b target, it achieves HMTD pass@1 of 0.61, while PRBO and Direct RL score 0.00. This judge-scored, single-behavior result is limited to the tested setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.