Exploring Agentic Schemes for Authoring Interactive Multiview Visualizations in Genomics

Abstract

Authoring genomics visualizations is challenging due to heterogeneous data, domain-specific constraints, and the gap between user intent and visualization expertise. Agentic approaches are widely assumed to help, but it is unclear which configuration is beneficial. We take a two-stage approach: we first evaluate vanilla LLM generation with Gosling across 159 cases, then compare four agentic schemes with varying numbers of specialist and reviewer agents against two baselines, direct generation and a fixed pipeline. The pipeline underperforms direct generation, and neither specialist agents nor a reviewer improves on a single iterative agent, which matches the most elaborate configurations at the lowest cost among them. Our results suggest domain-specific grammars and flexible agents are complementary: the grammar constrains the solution space while the agent searches within it.