A new research framework trains AI agents to sketch temporary interfaces instead of typing out walls of text.
Researchers propose GenUI-Harness, a two-agent system pairing a Tool Agent that retrieves information and executes tasks with a GUI Coder Agent that spots ambiguous requests and writes front-end code on the fly. Training the coder with reinforcement learning is tricky: testing whether a generated UI actually works requires running it, and letting another AI judge the output invites reward hacking. The team built two fixes - a sandbox called Dynamic UX for fast testing, and a Reward Auditor that catches judging patterns and turns them into a shared scoring rubric. They also introduce UI-TAU Bench, a 10-domain benchmark with 300 and 1,000-task test sets, to measure how well generated interfaces help users finish real tasks.
The numbers are the real story. Training with GenUI-Harness took a 4-billion-parameter model from completing 9.33% of tasks to 58% on the harder benchmark split, and a reviewer survey found generated UIs cut the average back-and-forth from 3.4 dialogue rounds down to 1.2. That is a meaningful case for structured interfaces over endless chat turns, especially for tasks like filling out forms or comparing database records where plain text becomes its own obstacle.
One claim is worth a raised eyebrow: the paper says its 4B model beats "Claude Opus 5" at 46.67%. No such model exists in Anthropic's current lineup, which tops out at Opus 4.8 - so that particular comparison can't be checked against anything real, and should be read as the paper's own framing rather than a verified fact.