← News

Agents & MCPSources: arXiv · 19 Nov 2026

StatePortal synthesizes checkpoints to reach deep states in GUI software testing

An arXiv paper describes an agent that builds replayable state checkpoints for interactive apps, plus a 61-task benchmark, with results reported by the authors.

In plain words

Testing apps with graphical interfaces is slow when a tester must click through many steps just to reach a rarely used internal state. StatePortal is an agent that creates a checkpoint, a saved snapshot of that state, which can be stored, restored and replayed later. The authors report it lets a GUI-testing agent reach such states far more often than without it.

What was announced

The paper argues that reaching a deep state, one requiring a long interaction history, is a recurring cost across code commits, and that GUI automation often assumes a DOM or accessibility tree that self-rendered engines do not expose. StatePortal takes a natural-language functional test case and read-only access to the target's source, then synthesizes a checkpoint that restores the state. An onboarding agent binds or writes a per-application hook once, without hand-written hook code, and checkpoints are accepted only when both a programmatic-state oracle and a visual oracle pass, with the visual oracle calibrated against a blinded cross-VLM reference.

The authors report that a GUI agent seeded with synthesized checkpoints completed 10.4% of trials versus 0% without seeding under the same pixel action space, while StatePortal, which also delivers the target action through the hook, completed 27.5%. Across 1,361 replays of 112 checkpoints, 98.90% reproduced the original verdict with zero model calls, and the hybrid oracle detected 96.25% of controlled visual faults that programmatic-state checks alone missed. The work accompanies DeepStateBench, 61 deep-state tasks across 10 interactive applications spanning five language families and six rendering stacks, with source sizes from 874 to 7.2M lines. Code is open-sourced on GitHub.

What it does not show

  • All figures come from the paper's own abstract; no independent evaluation is available, and the excerpt covers only the abstract, not the full paper.
  • The comparison to 0% without seeding applies to the same pixel action space and the three registered models, so results may not generalize to other setups.

For security and IT teams

Access
StatePortal requires only read-only access to the target application's source code, and the code is open-sourced on GitHub.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.