Cua-S1-FORMS: A Small, Specialized AI Model for Form Automation

Cua-S1-FORMS is a specialized AI model engineered to automate precise decision-making within digital forms, offering a lightweight alternative to general-purpose large language models (LLMs).

Developed by Dillon and Francesco from Cua, the model focuses on what they term "System 1" type tasks—quick, intuitive decisions like whether to check a box, click a button, or use a pre-filled value. Unlike token-generating LLMs, Cua-S1-FORMS scores a set of predefined actions for each form element.

This compact model, weighing in at just 706k parameters and 2.8 MB, was specifically trained on synthetic data for form interactions. It predicts actions such as 'use the given value', 'CHECK', 'CLICK', or 'SKIP' for structured elements extracted from a document.

Cua claims significantly higher accuracy compared to larger models like hosted Jev for form-specific tasks, achieving 99.7% correctness for the whole decision set against Jev's 83.6%. Its local inference speed of 7-9 milliseconds per form score also dramatically outperforms the 260-280 milliseconds seen with hosted solutions.

The interesting question for us was: what happens if you give a model an interface of current context, and a set of possible choices, and you ask it to return a probability for each choice?

The project aims to fill the gap between rigid, brittle scripts and computationally expensive, general-purpose AI agents. It targets recurring decisions that are too variable for simple scripting but too narrow to justify calling a full LLM.

Cua has open-sourced its synthetic data generation, training, evaluation, and driver integration under an MIT license. This allows developers to integrate the specialized model into their own computer-use agents, specifically for form handling.

While the current release is focused solely on forms, the developers envision specialists like Cua-S1 handling specific, well-understood decisions passed down from a more general agent, creating a modular approach to complex automation.