InterWebCoder trains multi-step webpages with executable tests
Submitted October 20, InterWebCoder replaces screenshot or GUI-agent rewards with replayable test scripts and state-level visual similarity for interactive webpage generation.
Why it matters
A rendered webpage does not establish that clicks, input, navigation and state transitions work. Executable tests provide faster, inspectable feedback for interaction behavior.
A closer look
Its TEDS pipeline reconstructs self-contained webpages and tests from real interaction trajectories, then attributes failures to either page code or the test. The authors retain 1,736 verified samples from 2,058 tasks and combine test-ID-weighted SFT with GRPO over test pass rate and state-level visual similarity. They report final scores of 48.54 on IWR-Bench and 48.42 on Vision2Web with a 9B model; local tests reduce reward computation from 92.40–130.01 seconds to 2.08 seconds per sample.
Who this is for
Teams working on interactive webpage generation, UI-to-code, test-driven training and generative-UI evaluation
Original sources
Revisions & corrections (1)
— New original research discovered outside the existing source directory; full paper and submission record verified, with attributed results, explicit limitations and counterexamples.