Reproduce “TRAP: Tree-Guided Reinforcement-Augmented Probing for Automated LLM Jailbreaking” 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
Mohammad Reza Mirbagheri, Amir Mahdi Sadeghzadeh
TRAP organizes jailbreak strategies in a fixed hierarchy and learns which paths work for each target, carrying selection preferences across objectives. It combines this policy with retrieval to adapt selected tactics to each objective. On seven AdvBench targets, TRAP has the highest observed attack success rate on six; on DeepSeek-V3.2, it reaches 97.9% versus AutoDAN’s 25.8%, with method-specific query settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.