Task-Dependent Evaluation of LLM Output Homogenization: A Taxonomy-Guided Framework

Shomik Jain, Jack Lanchantin, Maximilian Nickel +4

Large language models often generate homogeneous outputs, but whether this is problematic depends on the specific task. For objective math tasks, responses may vary in terms of pro…