Stanford introduced a simple prompting trick that boosts LLM creativity without retraining. Adding a short verbalized-sampling cue makes the model surface multiple possibilities instead of collapsing into predictable answers caused by alignment and typicality bias. By asking for a distribution, not a single reply, the model recovers much of its lost diversity and generates richer outputs.
#LLM #AIResearch #GenerativeAI #PromptEngineering #MachineLearning #ModelAlignment
// RELATED_UPLOADS