I wonder if this is related to their announcement that they will start watermarking text? Their method adds a statistical biases to the generated text, which they state doesn’t work in certain situations (short text sequences being one of them). This could be their solution to ensure watermarks are added to more outputs.