Researchers use persistent homology to map LLM latent space changes under adversarial attacks

Researchers have developed a new method using persistent homology to analyze the internal workings of Large Language Models (LLMs). This technique characterizes how adversarial inputs alter the geometric and topological structure of LLM latent spaces. The study found that adversarial attacks consistently lead to a topological compression, simplifying the latent space and collapsing features into fewer, larger ones, regardless of model architecture or specific attack type. AI

影响 Introduces a novel topological approach to understand LLM vulnerabilities and internal representations.

排序理由 Academic paper published on arXiv detailing a new interpretability method for LLMs.

在 arXiv cs.LG 阅读 →

• arXiv
• Inés García-Redondo
• LLM
• Persistent Homology

• 论文
• 安全

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →