Improving AI Writing: Verbalized Sampling

This scenario explores how to make all AI narrators more creative—with science 🧪 Ask your favorite LLM for a joke about coffee. ☕ Open a new chat and ask again. You’ll probably get the same one. That’s mode collapse—when models keep choosing the most likely answer until every path feels the same. It happens because of human typicality bias. During A/B testing, we tend to reward familiar phrasing over odd but valid alternatives. This lack of diversity forces us to battle clichés and predictable tropes, burying the unique voices the AI could otherwise produce. Verbalized Sampling asks models to output an explicit probability distribution over responses ("Generate 5 coffee jokes with their corresponding probabilities") and then pick from the less likely but still valid ones—trading predictability for creativity. Verbalized Sampling increases diversity by 1.6-2.1× in creative writing. This scenario attempts to replicate those positive outcomes within the AI Dungeon environment. It only injects the instruction at the very end of the prompt so you can safely import an intro, plot essentials, and cards from any other scenario, then try it out. 📝Note, it may occasionally "show its work," but a retry should correct it. Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity https://arxiv.org/html/2510.01171v3 🔮 Source Code https://gitlab.com/xilmanaath-ai-dungeon/verbalized-sampling Disclaimer: The following is a fan-made adaptation of Verbalized Sampling research. No lab scientists were harmed in the making of this experiment. Please support the official release—err research 🧠💫