May 2026 release notes for Document Understanding in Automation Cloud.

Correction, May 19, 2026: These changes were released in January 2026 but were not included in the release notes at the time.

UiPath AI extraction and classification models are now part of the UiPath Helix model family. The following models are available:

For more details, check the UiPath Helix Extractor 1.0 and UiPath Helix Extractor 2.0 pages.

UiPath Helix Extractor 2.0 is now available in preview for tenants in Europe and the United States. Built on a multimodal architecture, the new model processes document text and visual layout together, supporting non-Latin script languages and improving table extraction accuracy.

For more details, check the UiPath Helix Extractor 2.0 page.

You now start model training manually using the new Start Training button, instead of relying on continuous or automatic training in the background.

Each trainable model has its own button: classifiers expose it from the Model Training status pill (Helix Classifier on the Split & Classify page; Legacy Classifier in the Recommendations area on the Build page), and extractors expose it on each document type's annotation page or from the same status pill across the application.

A new changes counter shows how many annotations or classifications have accumulated since the last training, and the status pill now surfaces Queued, Failed (with a Retry action), and Trained states — including last-training date, duration, and base model version.

Schema or base model changes no longer auto-trigger training; a warning is shown instead so you can decide when to start a new run. Zip imports continue to queue a training automatically.

For details on button states, the changes counter, and the training lifecycle, see Start a training run.