DashArena: New Benchmark for Interactive Analytic Dashboard Generation

Researchers have introduced a new benchmark called DashArena, aimed at evaluating large language models (LLMs) in their ability to create interactive analytic dashboards. The details are shared in a paper available on arXiv (ID: 2608.10567). This benchmark tackles the challenge of assessing dashboard generation, which is often open-ended. DashArena stands out as the first of its kind for task-grounded and open-ended dashboard creation. It requires models to produce a dashboard along with a replayable interaction trajectory. A browser executor then replays this path, providing visual proof of the analytic process. A vision-language model (VLM) judge assesses the outputs, and a leaderboard is created using Bradley–Terry aggregation. The researchers also developed the DashJudge-8B model based on the judge. You can read the paper at https://arxiv.org/abs/2608.10567.