REFLECTOR: Internalizing Step-wise Reflection against Indirect Jailbreak
The paper introduces Reflector, a framework designed to enhance the safety of Large Language Models (LLMs) against sophisticated jailbreak attacks. It employs a two-stage approach that combines teacher-guided generation and reinforcement learning to foster self-reflection capabilities. Empirical results indicate that Reflector significantly improves defense success rates and overall performance on various benchmarks.
- ▪Reflector addresses vulnerabilities in Large Language Models by internalizing self-reflection.
- ▪The framework achieves over 90% defense success rates against complex indirect attacks.
- ▪Reflector enhances both task-specific and general utility, yielding a 5.85% gain on GSM8K.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.20654 (cs) [Submitted on 20 May 2026] Title:REFLECTOR: Internalizing Step-wise Reflection against Indirect Jailbreak Authors:Jiachen Ma, Jiawen Zhang, Xiangtian Li, Bo Zou, Chaochao Lu, Chao Yang View a PDF of the paper titled REFLECTOR: Internalizing Step-wise Reflection against Indirect Jailbreak, by Jiachen Ma and 5 other authors View PDF HTML (experimental) Abstract:While Large Language Models (LLMs) demonstrate remarkable capabilities, they remain susceptible to sophisticated, multi-step jailbreak attacks that circumvent conventional surface-level safety alignment by exploiting the internal generation process.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
