Interactive data tools are moving toward a more connected way of working, where a change in one view can affect other charts, dashboards, or application states. Two publications dated 19 October 2026 and October 22, 2026 focus on that shift from different angles: one presents Flint-interactive, a language for building interactive data visualizations, while the other examines how vision-language models understand those visualizations.
Together, the work points to a clear challenge. People want AI agents to create visualizations that help users explore and understand data, but producing high-quality interactive charts remains difficult. The tools must support more than a static image or a basic chart. They need to respond to user actions, connect different views, and present information in ways that people can follow.
A language for connected visual experiences
Flint-interactive is designed to make interactive data visualizations easier to create. It supports dashboards, data stories, and interactive artifacts, giving users a way to build experiences that connect multiple views instead of treating every chart as a separate object.
Its main strength comes from the Flint compiler. The compiler translates user interactions into semantic events, then uses high-level operators to update visualizations. In plain terms, an action made in one part of an experience can carry meaning that other parts understand, allowing the system to update the right visualizations or application states.
That structure supports interactive data stories, dashboards, and connected experiences. A user’s action in one view can update another visualization, or change an application state tied to the broader experience. The result is a model for interaction that goes beyond the built-in behaviors found in many visualization libraries.
Those built-in interactions can help with common chart tasks, but they are not expressive enough for every bespoke interactive chart. When a project needs a specific relationship between user actions and visual updates, developers need more than a fixed collection of controls. Flint-interactive addresses that gap by providing a language and compiler-based system for describing those relationships.
Why this matters for AI-generated visualizations
AI agents can generate interactive visualizations to help users explore and understand their data. That possibility brings a practical problem: creating a high-quality interactive visualization has high generation costs, and reliability issues can make the result difficult to trust or use.
Flint-interactive offers a way to think about that problem through structured interactions. Instead of requiring every behavior to be built as a separate, narrowly defined feature, the system translates interactions into semantic events and applies high-level operators to update the visualization. This gives AI agents a clearer foundation for generating connected experiences, at least in the language’s design.
The need for that foundation becomes clearer when considering the second publication. Models performed worse than humans on average when understanding data visualizations, and the gap remained even under lenient criteria. That finding matters because an AI agent that creates a visualization must also handle the meaning carried by charts and the relationships between their parts.
All models produced error patterns that were reliably distinct from those of humans. The point is not only that models made mistakes. Their mistakes followed patterns that differed from human errors, showing that model understanding of visualizations does not work in the same way as human understanding.
Building tools while measuring understanding
The two lines of work fit together without solving the same problem. Flint-interactive focuses on how to create interactive data visualizations, while the model evaluation focuses on how well vision-language models understand those visualizations. Better creation tools do not remove the need to measure whether models interpret charts correctly.
That distinction is important for AI agents. An agent may generate a dashboard or data story, yet still struggle to understand what the visualization communicates. It may also face the separate task of connecting user actions to the correct visual updates. Flint-interactive targets the interaction layer, while the evaluation findings highlight limits in visual understanding.
The work comes from researchers and practitioners including Chenglong Wang, Principal Researcher; Yuwei Xiao, Computer Science Ph.D. Student; Alper Sarikaya, Senior Data Visualization Engineer; Michel Galley, Senior Principal Research Manager; Chris Quirk, Partner Researcher; and Jianfeng Gao, Technical Fellow & Corporate Vice President. The visualization-understanding research includes Arnav Verma, Kushin Mukherjee, Christopher Potts, Elisa Kreiss, and Judith E. Fan, whose affiliations include Stanford University, MIT, and UCLA.
The larger message is simple: interactive visualization needs both stronger construction tools and careful testing of model understanding. Flint-interactive provides a language for building dashboards, data stories, and connected artifacts. The evaluation results show why systems that work with those visualizations still need close attention to reliability and interpretation.
Based on
