Language-Models

The earth is flat and the sun is not a star: The susceptibility of GPT-2 to universal adversarial triggers

Examines universal adversarial triggers in natural language models, showing how specific text sequences can manipulate GPT-2's outputs on …