FlowEval asks whether an AI-generated interface works, not only whether it looks good

On May 5, 2026, a team involving Apple and academic collaborators introduced FlowEval for evaluating AI-generated interfaces. An interface created by a coding agent can have convincing buttons, cards and typography, but that does not show that a person can complete the task it was built for.

The proposal does not ask another model to look at a screenshot and assign a score. It starts with a reference website considered functional, runs concrete tasks on it and on a generated version, and compares the traces of both journeys. If the steps, screens and transitions in the generated interface resemble those in the reference, its score rises.

Measuring a flow, not only a screen

FlowEval uses a computer-use agent to open the reference site and the generated analogue, move the pointer, type, click and capture screens while carrying out a task. Each journey becomes a sequence of visual representations. The system then calculates how similar those sequences are using metrics borrowed from other areas of machine learning.

One is dynamic time warping, which can compare journeys of different lengths. Another searches for overlap between subsequences. The best-performing measure in the study was Word Mover’s Distance, a distribution-distance measure applied here to representations of screen states.

The practical goal matters. A company can choose reference designs and tasks that express its own product language. Rather than asking an automatic judge whether a screen looks good in the abstract, the method checks whether the result supports a flow the team has already validated.

What the authors found

To test whether the metric approached human judgment, the researchers generated 945 interfaces from 27 reference sites, seven code-generation models and five repeats per prompt. They then collected 429 blind pairwise comparisons between generated websites. Two human-computer-interaction experts tested the interfaces and chose which worked better.

According to the preprint, Word Mover’s Distance reached a Spearman correlation of 0.96 with those experts’ ranking and 73.0% binary agreement. It also outperformed a multimodal language-model judge in that experiment. That is encouraging because it produces an interpretable score: a team can inspect where the journeys diverged.

The figures need to be read as a feasibility demonstration. Validation involved only two annotators and one navigation agent. The models also generated pages in a one-shot setting with structured requirements, which is clearer than many real requests to a coding assistant. The authors acknowledge that agents and APIs change quickly, and that a weak navigation agent can fail even when an interface is implemented well.

One part of evaluation, not the complete verdict

Interface quality does not fit in one journey. Accessibility, performance, security, aesthetics, content comprehension and user satisfaction still require other tests. A highly original interface may also depart from a reference for a good reason, and public benchmark flows do not cover every case involving logins, payments or complex state.

For that reason, FlowEval is more useful as a complement than a replacement for human review. For teams asking coding agents to make interfaces, it offers one specific automatable question: does this version let a person do what a good reference lets them do? Turning that question into traces and comparing them can expose failures that a polished screenshot keeps hidden.