- Who it helps
- Researchers comparing computer-use agents and developers designing diagnostic evaluations.
- Availability
- The October 14 paper is public. Its abstract promises code and data; a usable release was not verified in this check.
Read the route through the task
The authors describe tasks across ten applications and a graph of states and transitions. Their annotations distinguish progress from repeated or off-path actions. This makes a completed task inspectable as a sequence, so an evaluator can ask how much unnecessary work or recovery it required.
Keep the conclusion within the benchmark
The reported analysis finds deviations or loops within many successful runs. It motivates checking trajectories alongside outcomes. It does not prove recovery under every external change or provide a deployed monitor. A useful independent test would introduce a known disturbance and check whether the agent recovers without repeating an unsafe action.
What you can do
- Compare the paper’s task graph and error labels with one GUI task you already know how to verify. Open reference ↗
Where the evidence stops
Only the abstract and publication record were checked. No benchmark or agent was run; external-change recovery remains partly unresolved.
Sources read. Sources checked October 22, 2026 UTC. No external software workflow, model inference or reported experiment was executed.
Sources & further reading
Paper · paper
GraceBench: evaluating GUI-agent trajectories ↗
Original abstract read for trajectory deviations, loops and recovery. The promised code/data release was not verified.
GraceBench authors · page checked 2026-10-22 · Published 2026-10-14
Coverage note: The original abstract was read. Public code/data availability, reproduced trajectories and independent discussion were not established.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
GraceBench evaluates the path a GUI agent takes, including deviations, loops and recovery that a final success score can hide.