Reproduce “META-IR: Meta-Introspective Reasoning for Adaptive Safety Alignment in LLM” 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
Mahabubur Rahman Miraj, Hongyu Huang, Nankun Mu, Chao Chen, Shanqin Yi
META-IR addresses a weakness in language-model safety by structuring safety analysis, searching and pruning reasoning paths, and adapting that policy across jailbreak families. On LLaMA-3.1-8B-Instruct, it reports Adv of 0.089 and SVR of 0.061, compared with STAIR-SFT at 0.281 and 0.211. Its WJ-B score is lower, while MGS and MTBench are higher. These benchmark results do not establish protection against every jailbreak.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.