Reproduce “REFLECTOR: Internalizing Step-wise Reflection against Indirect Jailbreak” 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
Jiachen Ma, Jiawen Zhang, Xiangtian Li, Bo Zou, Chaochao Lu, Chao Yang
The REFLECTOR framework internalizes step-wise reflection within large language models to provide robust defense against indirect jailbreak attacks, notably boosting defense success rates by over 80 percentage points against various attack types. This approach simultaneously enhances general utility, such as achieving a 5.65% absolute gain on mathematical reasoning benchmarks.
Key insights
- Internalizing self-reflection directly into an LLM's generation trajectory offers a robust defense against indirect jailbreaks by enabling continuous, dynamic vigilance and correction.
- A two-stage training methodology (SFT for capability injection and RL for optimization) is crucial for stable and effective implementation of intrinsic reflection.
- Enhancing safety through internalized reflection can unexpectedly boost general utility and reasoning capabilities, challenging the conventional 'alignment tax' by making safety and general intelligence mutually reinforcing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.