DataMagic: Authoring Data Videos through Declarative Multi-Agent Orchestration

DataMagic. Authoring Data Videos through Declarative Multi-Agent Orchestration. By Yupeng Xie and five others, from the Hong Kong University of Science and Technology, Guangzhou. Published on the twenty-seventh of September. Its code repository is worked out: about one hundred and ninety source files, with a readme and a licence.

Data videos tell a story with data: dynamic charts, voice narration, and animations kept in sync. They have become a widely adopted form of data storytelling, but making one asks for three kinds of expertise at once. Data analysis, narrative design, and video editing. The existing tools each cover part of that. Static visualization tools have no narrative or animation. Authoring tools start from charts someone has already prepared, not from raw data. Pixel-level models generate a whole video end to end, but cannot guarantee that the data shown is accurate or traceable to its source.

Generating these videos automatically, from start to finish, runs into two core challenges. One is representation: how to describe charts, narration and animation together, along with how they line up in time. The other is search: how to move efficiently through a vast space of designs to find compositions whose narrative holds together.

DataMagic authors data videos straight from raw tabular data, through what the authors call declarative multi-agent orchestration. Each of the two challenges gets its own answer.

For representation, there is DVSpec, a declarative specification. It unifies charts, narration and animations, using references bound to the data and synchronization stated declaratively. That is what secures data provenance, and it makes audio-visual alignment automatic. For search, there is a strategy called Generate-then-Orchestrate. Multiple agents generate candidate scenes in parallel, and a global orchestration step then optimizes how coherent the narrative is across them. DVSpec also acts as shared state for three complementary modes of interaction, which lets the system range from full automation to fine-grained human control.

Evaluation uses one hundred and nine real-world samples. Even the most advanced LLM, with GPT-5 as the example given, scores only two point one three out of five, with execution success rates between forty-eight point six two and eighty-six point two four percent.

DataMagic raises quality to three point eight nine, an improvement of eighty-three percent, with success rates above ninety-five percent. The largest gains are in the animation and narrative dimensions. A user study compares it with a conversational workflow built on an LLM. DataMagic cuts task time by seventy-nine point seven percent, a large gain in creation efficiency, and lowers the cognitive load users report.

Everything here hangs on the specification. Once charts, words and motion are written in one declarative form that points back at the data, accuracy and timing stop being things the generator has to get lucky with. They become properties of the document, and the agents are left to search for the story.