Beyond One Output: Visualizing and Comparing Distributions of Language Model Generations

arXiv:2604.18724v1 Announce Type: new Abstract: Users typically interact with and evaluate language models via single outputs, but each output is just one sample from a broad distribution of possible completions.