Macaron-A2UI: A Model for Generative UI in Personal Agents

Researchers introduced Macaron-A2UI, a Generative UI model for personal agents that can generate both language and executable interface actions.

They trained 30B, 235B, and 754B models on a large Generative UI corpus and evaluated them on A2UI-Bench.

The system reached a top score of 75.6 without explicit schema hints, outperforming prior schema-based baselines.

The team also released the models, benchmark, and evaluation protocol to support further research on interactive assistants.