Adobe's EditAgent shows an AI judge picks better when you hand it measurements, not just the finished output

If your AI checker keeps approving bad output, the problem may be what you show it, not which model you picked.

What happened

Adobe Research and UC San Diego published EditAgent, an image editing agent that works more like a Photoshop professional. It splits a request into steps, tries the general purpose editor alongside specialized tools for each step, and lets a standard language model pick the best result.

In a test of 600 edits blind rated by five professionals, the general purpose editor Flux-2-Klein-9B failed to produce a usable result 42.7% of the time. Specialized tools solved roughly two thirds of those failures, and more retries with the same editor barely helped. Professionals preferred variants generated in parallel from the original over step by step refinement in 78.5% of decided comparisons, with half as many evaluation passes.

The key finding is the verification record: a close up of the edited area, a map of changed regions, and measured brightness and color statistics. Without it, none of the seven language models tested matched the best model fine tuned for judging. With it, all seven beat every fine tuned model. Qwen3.8-27B made a usable pick 81.3% of the time against 74.8% for EditScore-72B. The agent also checks whether later edits undid earlier ones. The tradeoff is about ten times the compute of a single editor.

My take

Swap images for documents and this is the fix for most weak AI review steps I see:

1. Do not ask a model "is this good?" Give it facts it cannot see by reading: word count, which CRM fields changed, whether required fields are filled, a readability score, a diff against the source.
2. Generate two or three candidates from the original input and compare them in one call, instead of revising the same draft over and over.
3. After the last step, check that earlier requirements still hold. Rewrites love to undo the compliance line.

Measured checks are cheap code. They make an ordinary model a reliable reviewer.