Reproduce “Beyond the Tip of Efficiency: Uncovering the Submerged Threats of Jailbreak Attacks in Small Language Models” 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

Sibo Yi, Tianshuo Cong, Xinlei He, Qi Li, Jiaxing Song

An empirical study systematically assessed the security performance of 13 state-of-the-art Small Language Models (SLMs) against various direct and sophisticated jailbreak attacks. The work found SLMs generally more susceptible to advanced attacks than larger models, often displaying "false robustness" from limited generalization, while demonstrating the effectiveness of existing prompt-level defenses.

Key insights

  • Some SLMs exhibit "false robustness" against certain jailbreak attacks, where a lack of generalization capabilities, rather than strong defenses, leads to meaningless, harmless outputs.
  • Parameter size, architectural compression (e.g., parameter sharing), and biased knowledge distillation negatively correlate with SLM security, degrading their defense capabilities.
  • Quantization techniques (e.g., AWQ, GPTQ-Int4/Int8) applied to SLMs surprisingly do not significantly weaken security and can, in some cases, slightly enhance robustness.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.