General-purpose text decoding approaches are usually adopted for dialogue
response generation. Although the quality of the generated responses can be
improved with dialogue-specific encoding methods, conversational decoding
methods are still under-explored. Inspired by \citet{wu2023learning} that a
good dialogue feature space should follow the rules of local
研究人类在对话中遵循 Uniform Information Density 原则的情况下，使用 GPT-2 在 Persona-Chat 数据集上生成响应，发现解码算法促进 Uniform Information Density 并不会生成更高质量的响应，相反，鼓励非一致性响应则是解决质量退化问题的潜在解决方案。