This and related techniques are trivially foolable by fine-tuning the model. They're also trivially foolable by using sampling techniques or settings which enc

This and related techniques are trivially foolable by fine-tuning the model. They're also trivially foolable by using sampling techniques or settings which encourage the model to generate rare words a lot. Also foolable with filter-assisted decoding: https://paperswithcode.com/paper/most-language-models-can-be-poets-too-an-ai https://paperswithcode.com/paper/most-language-models-can-be...