Abstract
Recently, generative pre-trained models have made significant strides,particularly highlighted by the release of ChatGPT and GPT-4, which exhibitsuperior cross-domain capabilities. However, these models still face challengeson constrained writing tasks like poem generation under open-domain titles. Inresponse to this challenge, we introduce Block Inverse Prompting (BIPro)constrained generation framework. BIPro leverages two block inverse promptingmethods, revise and rewrite, that mimic the process of human text writing usingblock generative models. It significantly improves the zero-shot generationquality on the formidable constrained generation task of open-domaintraditional-form Chinese poem generation. Based on a less powerful blockgenerative model GLM-10B-Chinese, poems composed via BIPro without priming oradditional training outperform both most advanced direct generative systemslike GPT-4 or GLM-4 and best domain-specific systems such as Yusheng,Shisanbai, or Baidu Poetry Helper in human evaluation by proficient poets.Finally, BIPro considerably narrows the gap between AI-generated works andshort-listed human literary arts in another human evaluation, unveiling thepromising potential of block generative models in improving the quality ofconstrained generation.