Reproduce “Rethinking LLM Jailbreak Evaluation: Continuous Risk Landscapes Beyond Binary Outcomes” 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
Yukai Zhou, Wenjie Wang
The authors estimate jailbreak risk across alternative first-token branches by fixing each candidate opening, sampling continuations, and scoring them with a frozen judge. They compare attack selection by ordinary policy risk with selection by risk after a controlled decoder rerouting intervention. On Gemma 3, selection reverses on held-out DHS within the tested nine-attack portfolio, not generally.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.