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

A team from the University of Washington and the Allen Institute for AI has published a paper introducing GROVE, a novel interactive visualization tool designed to address a critical blind spot in how we interact with large language models (LLMs). Currently, users typically see and evaluate a single output from models like GPT-4 or Llama 3, but each response is just one sample from a vast distribution of possible completions. This hides crucial structural information such as multiple valid answer modes, uncommon edge cases, and sensitivity to minor prompt tweaks, leading developers and researchers to over-generalize from anecdotes. Informed by a formative study with 13 AI researchers, GROVE was built to visualize these distributions by representing multiple LM generations as overlapping paths through a dynamic text graph, making shared phrases and divergent branches immediately apparent.

GROVE enables a hybrid workflow for prompt engineering and model evaluation. The tool’s graph-based summaries excel at helping users make structural judgments, such as assessing the overall diversity and clustering of possible model responses to a given prompt. However, the team’s three crowdsourced user studies (with 47, 44, and 40 participants) found that for detail-oriented tasks, directly inspecting raw text outputs remains superior. This suggests GROVE is best used as a complementary lens, not a replacement. By making the inherent stochasticity and range of LLMs visually tangible, it provides a more systematic foundation for debugging prompts, comparing model behaviors, and understanding when a single output is representative or merely a lucky—or unlucky—sample.