Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

X-Teaming Evolutionary M2S: Automated Discovery of Multi-turn to Single-turn Jailbreak Templates

Created by
  • Haebom

Author

Hyunjun Kim, Junwoo Ha, Sangyoon Yu, Haon Park

Outline

This paper presents research on a multi-turn-to-single-turn (M2S) technique that condenses repetitive red team activities into a single, structured prompt. Unlike previous studies that rely on a few handwritten templates, this paper proposes the X-Teaming Evolutionary M2S framework, which automatically discovers and optimizes M2S templates using a language model (LLM)-based evolutionary algorithm. It uses smart sampling from 12 sources and a StrongREJECT-inspired LLM as a judge, resulting in a fully auditable log. After five evolutionary generations, with a success threshold of 0.70, we achieve an overall success rate of 44.8% (103 out of 230) on two new template families and GPT-4.1. Through 2,500 cross-model evaluations, we demonstrate that structural improvements are transferable, but vary across target models. We find a positive correlation between prompt length and scores, highlighting the importance of length-sensitive judgement. The source code, configuration, and results are available on GitHub.

Takeaways, Limitations

•
Takeaways:
◦
We present a novel framework for automatically generating and optimizing M2S templates using a language model-based evolutionary algorithm.
◦
Emphasizes the importance of threshold setting and cross-model evaluation for successful M2S template generation.
◦
Suggesting future research directions by revealing the correlation between prompt length and performance.
◦
Although it shows the possibility of structural improvement, it suggests that performance differences between models should be taken into account.
•
Limitations:
◦
The success rate of 44.8% still leaves room for improvement.
◦
Performance is poor for certain models (both models score 0 at the same threshold).
◦
There is a dependency on the LLM used (GPT-4.1), and further research is needed on generalizability to other LLMs.
◦
Further analysis and in-depth understanding of the correlation between prompt length and performance is needed.
👍
Made with Slashpage