Generating Multiple Options Boosts LLM Answer Diversity, Study Finds

Stanford researchers found LLM answer convergence stems from annotators' familiarity bias in RLHF, and simply prompting models to generate multiple options boosts diversity 2.1x without retraining. Critics caution that models lack introspective access to true probabilities, and forcing more options than actually exist can induce hallucination.

2026-08-23 ~ 2026-08-23 · 3 related posts