Reproduce “Robustifying Safety-Aligned Large Language Models through Clean Data Curation” 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

Xiaoqun Liu, Jiacheng Liang, Muchao Ye, Zhaohan Xi

A framework called CTRL (Clean DaTa CuRation for LLMs) enhances the safety alignment of large language models by curating general-domain clean texts based on their perplexity. This method mitigates training-based jailbreaking attacks, demonstrating reduced attack success rates and harmfulness scores while preserving or improving text quality.

Key insights

  • Safe responses to security-sensitive queries generally exhibit lower perplexity for LLMs compared to harmful or even general-domain responses.
  • Curating general-domain clean texts to reduce their perplexity can proactively enhance LLM safety alignment without requiring specific knowledge of adversarial attacks.
  • An iterative text revision process using a combination of output sampling and quality-controlled beam search can simultaneously reduce perplexity and maintain or improve text quality (readability and helpfulness).

Summary by alphaXiv.

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