Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity

It turns out ChatGPT is just as cloying in Chinese as it is in English, owing to some bizarre tics in LLM training data.