À propos
October 2026 paper pairing database tool execution with task-specific frontend generation. GenUI-Harness adds sandboxed interaction rollouts and reward auditing, while UI-TAU Bench tests whether generated interfaces help agents collect input and complete database-backed tasks across ten domains.
Résumé
Why GenUI-Harness Adds New Signal:
1. Grounded interaction loop: Separates retrieval and execution from generated controls that clarify requests and gather structured input
2. Executable evaluation: UI-TAU Bench offers 300-task Lite and 1,000-task Full splits; reported main comparisons use Lite
3. Training and interaction evidence: A trained 4B model reaches 58% Pass@3; a selected 100-task reviewer study reduces dialogue from 3.4 to 1.2 rounds
4. Evaluation limits: The browser user simulator sees source code and reference goals, and the reviewer comparison selects consistently usable interfaces; results do not establish general usability
1. Grounded interaction loop: Separates retrieval and execution from generated controls that clarify requests and gather structured input
2. Executable evaluation: UI-TAU Bench offers 300-task Lite and 1,000-task Full splits; reported main comparisons use Lite
3. Training and interaction evidence: A trained 4B model reaches 58% Pass@3; a selected 100-task reviewer study reduces dialogue from 3.4 to 1.2 rounds
4. Evaluation limits: The browser user simulator sees source code and reference goals, and the reviewer comparison selects consistently usable interfaces; results do not establish general usability
Tags
generative-uidata-grounded-uiui-tau-benchreinforcement-learninginteractive-agents