CUA releases open-source CUA-S1-FORMS for bounded web-form actions

• CUA has open-sourced CUA-S1-FORMS, the first member of its small, task-specialized System One model family for computer use, in the announcement.
• The model chooses only from a bounded form vocabulary, FILL, CHECK, CLICK, or SKIP, while form values come from a supplied document, per the action contract.
• A local application validates the choices and orders field fills, checkboxes, and an authorized submit action, as the action contract describes.
• The MIT release includes synthetic-data generation, training, evaluation, and Cua Driver integration in the source-release thread, and CUA is targeting repetitive, bounded, high-volume workflows in the deployment scope note.

The model card puts the model at roughly 706,000 parameters, using a byte-level two-layer Transformer with four attention heads. Its shuffled-context control fell to 37% top-1 accuracy, versus 99.95% on the form-disjoint synthetic test, a compact check that the scorer is reading the UI context rather than exploiting option frequency.

CUA-S1-FORMS is a one-pass option scorer for GUI form filling, not a text generator. The model card says it receives one UI element plus typed candidate options and returns a probability for each option.

That makes form filling a selection problem. The model code, synthetic-data generator, evaluation scripts, and live Driver loop are published in the CUA-S1 source tree, inside Cua's MIT-licensed repository.

Each actionable form element is scored independently, and the model can score those elements in parallel in one batch. A fill choice is represented by one candidate for each entity extracted from the supplied document; CHECK, CLICK, and SKIP are additional fixed candidates in the model card.

The action vocabulary is deliberately small:

• FILL: associate the current control with a document value.
• CHECK: select a checkbox or equivalent control.
• CLICK: activate a candidate control.
• SKIP: leave the current element alone.

The model returns probabilities, then downstream code selects from the live candidate set. No free-form answer has to be parsed back into an action.

The planner, rather than the model, owns sequencing. CUA says the application validates candidate choices, fills fields first, handles checkboxes next, and submits only when authorized in the action contract.

Cua Driver executes the prebuilt actions. The model card assigns it the rest of the loop: snapshot the form, score choices, order the plan, execute it, and observe fresh state. The separation keeps the scorer focused on a local choice while the driver handles UI control.

CUA generated 10,000 synthetic training episodes from forms containing two to 16 fields drawn from a 55-concept catalog, according to the model card. The generator adds distractors, confusable concepts such as email versus street, title variations, and disjoint train and test field signatures.

The published architecture uses:

• Byte-level embeddings for UI context and option text.
• A two-layer, width-128 Transformer encoder with four heads for context.
• Option-as-query attention over context tokens.
• A shared dot-product head and softmax over the currently live options.

The resulting checkpoint is about 2.8 MB. CUA trained it with AdamW, cosine decay and warmup for six epochs, using cross-entropy over the candidate set, per the model card.

CUA's model card reports 99.95% top-1 accuracy on its synthetic form-disjoint test. Its live demo evaluation reached 100% over 196 decisions from three real forms and three PDFs.

The same card records the shuffled-context control at 37%, and lists the system's operating constraints:

• Values must arrive as label-value pairs extracted from the supplied document.
• The scorer cannot invent missing values.
• Training is synthetic, while the real evaluation set contains six form-document combinations.
• Its vocabulary is English-centric.

Those boundaries are central to the result: the release measures a tightly specified form-selection loop, not broad autonomous computer use.

CUA calls Forms the first specialist in the CUA-S1 family and is soliciting the next tasks from organizations running repetitive, bounded, high-volume computer workflows. The project is positioning its next model targets around the same pattern: a closed action set, supplied task data, and a driver that executes the selected plan.