keyword

prompt robustness

Prompt robustness is the ability of an artificial intelligence model, particularly a large language model, to maintain consistent, accurate, and reliable performance despite variations, noise, or adversarial alterations in its input prompts. While minor modifications in phrasing, word choice, syntax, or formatting can often cause significant and unintended shifts in model outputs, a prompt-robust model demonstrates resilience to such sensitivity. This encompasses stability against benign variations, such as natural paraphrasing, typographical errors, and multilingual translations, as well as defense against deliberate adversarial attacks, such as prompt injections and malicious perturbations designed to bypass safety guardrails or induce task failure. Evaluating and improving prompt robustness is essential for ensuring that language models behave predictably, safely, and effectively in diverse real-world applications.

1 item