Researchers introduce GROVE, an interactive visualization tool for exploring distributions of language model outputs

A team including researchers from the University of Washington and others has introduced GROVE, an interface designed to make the distributional properties of language model outputs visible and explorable. Rather than showing users one completion at a time, GROVE renders multiple generations as paths through a shared text graph, allowing patterns of divergence and convergence to emerge.

The work is grounded in empirical observation: a formative study with 13 LM-using researchers identified moments when stochasticity matters to practitioners, how they mentally model distributions over language, and failure points in current interaction paradigms. This informed the design of the visualization approach.

Three separate user studies evaluated GROVE against complementary tasks. Across 47, 44, and 40 crowdsourced participants, the tool showed measurable advantages in structural reasoning—particularly in assessing diversity and identifying common modes—while direct output browsing remained more efficient for questions requiring close textual examination. The findings suggest neither approach fully replaces the other.

The research contributes to a growing area of interest in human-centered tools for LM development: making implicit distributional knowledge explicit through interface design, and testing those designs rigorously with users performing realistic tasks.