Latent Undertow: How Ordinary Typos Break Probes

LLMs handle ordinary typing variation fluently: a typo or missing punctuation leaves both user intent and the model's response substantively unchanged.…

LLMs handle ordinary typing variation fluently: a typo or missing punctuation leaves both user intent and the model's response substantively unchanged. Yet probes that detect malicious prompts by reading the model's hidden states tell a different story: the same edit rotates the…

Read the original source — arxiv.org

paper · Shared by tscosj

0 comments

No comments yet.