VizPilot's MLLM browser extension auto-generates interactive visualization onboarding

Composite visualizations—multi-panel charts that combine several views—can overwhelm novice users despite their power for representing complex datasets. Existing onboarding methods are clunky: they're platform-dependent, demand heavy manual authoring, and often fail on structurally intricate designs. VizPilot, created by Nishaanthini Gnanavel and Yong Wang, flips this by reverse-engineering the structure of raw visualization artifacts to build interactive onboarding experiences automatically.

VizPilot runs as a browser extension and relies on a two-stage pipeline in its Composite Visualization Analyzer. First, multimodal LLMs decompose a visualization into distinct visual components, then extract structured explanations and map each component to precise SVG elements for reliable highlighting and interaction. The Onboarding Interface offers two modes: a guided narrative scrollytelling experience and free exploration, letting users learn progressively or on demand. Developers only need to supply a short visualization description and optional interaction source code. The authors validated VizPilot via comparative input-modality analysis, a usage scenario demonstrating reduced authoring effort, and a user study measuring cognitive load. Results indicate it effectively automates onboarding authoring while making composite visualizations more usable and accessible. Accepted at IEEE VIS 2026, VizPilot points toward a future where complex charts teach themselves.