Talking Nonsense: Probing Large Language Models' Understanding of Adversarial Gibberish Inputs
This work addresses the problem of LLM robustness and alignment for researchers and developers, revealing vulnerabilities in handling nonsensical inputs, though it is incremental in exploring adversarial probing.
The study investigated whether large language models (LLMs) can generate coherent responses from adversarial gibberish inputs, finding that manipulation efficiency depends on target text length and perplexity, and that generating harmful texts is not more difficult than benign ones, indicating alignment issues for out-of-distribution prompts.
Large language models (LLMs) exhibit excellent ability to understand human languages, but do they also understand their own language that appears gibberish to us? In this work we delve into this question, aiming to uncover the mechanisms underlying such behavior in LLMs. We employ the Greedy Coordinate Gradient optimizer to craft prompts that compel LLMs to generate coherent responses from seemingly nonsensical inputs. We call these inputs LM Babel and this work systematically studies the behavior of LLMs manipulated by these prompts. We find that the manipulation efficiency depends on the target text's length and perplexity, with the Babel prompts often located in lower loss minima compared to natural prompts. We further examine the structure of the Babel prompts and evaluate their robustness. Notably, we find that guiding the model to generate harmful texts is not more difficult than into generating benign texts, suggesting lack of alignment for out-of-distribution prompts.