OpenCUA’s open supply computer-use brokers rival proprietary fashions from OpenAI and Anthropic

A brand new framework from researchers at The College of Hong Kong (HKU) and collaborating establishments offers an open supply basis for creating strong AI brokers that may function computer systems. The framework, known as OpenCUA, consists of the instruments, knowledge, and recipes for scaling the event of computer-use brokers (CUAs).

Fashions skilled utilizing this framework carry out strongly on CUA benchmarks, outperforming present open supply fashions and competing carefully with closed brokers from main AI labs like OpenAI and Anthropic.

The problem of constructing computer-use brokers

Laptop-use brokers are designed to autonomously full duties on a pc, from navigating web sites to working complicated software program. They will additionally assist automate workflows within the enterprise. Nonetheless, essentially the most succesful CUA techniques are proprietary, with vital particulars about their coaching knowledge, architectures, and improvement processes saved non-public.

“As the shortage of transparency limits technical developments and raises security issues, the analysis group wants actually open CUA frameworks to check their capabilities, limitations, and dangers,” the researchers state in their paper.

On the identical time, open supply efforts face their very own set of hurdles. There was no scalable infrastructure for gathering the varied, large-scale knowledge wanted to coach these brokers. Current open supply datasets for graphical person interfaces (GUIs) have restricted knowledge, and plenty of analysis tasks present inadequate element about their strategies, making it tough for others to duplicate their work.

Based on the paper, “These limitations collectively hinder advances in general-purpose CUAs and limit a significant exploration of their scalability, generalizability, and potential studying approaches.”

Introducing OpenCUA

OpenCUA is an open supply framework designed to handle these challenges by scaling each the info assortment and the fashions themselves. At its core is the AgentNet Instrument for recording human demonstrations of laptop duties on totally different working techniques.

The device streamlines knowledge assortment by working within the background on an annotator’s private laptop, capturing display screen movies, mouse and keyboard inputs, and the underlying accessibility tree, which offers structured details about on-screen parts. This uncooked knowledge is then processed into “state-action trajectories,” pairing a screenshot of the pc (the state) with the person’s corresponding motion (a click on, key press, and so forth.). Annotators can then assessment, edit, and submit these demonstrations.

Utilizing this device, the researchers collected the AgentNet dataset, which incorporates over 22,600 job demonstrations throughout Home windows, macOS, and Ubuntu, spanning greater than 200 purposes and web sites. “This dataset authentically captures the complexity of human behaviors and environmental dynamics from customers’ private computing environments,” the paper notes.

Recognizing that screen-recording instruments increase important knowledge privateness issues for enterprises, the researchers designed the AgentNet Instrument with safety in thoughts. Xinyuan Wang, co-author of the paper and PhD scholar at HKU, defined that they applied a multi-layer privateness safety framework. “First, annotators themselves can totally observe the info they generate… earlier than deciding whether or not to submit it,” he advised VentureBeat. The info then undergoes guide verification for privateness points and automatic scanning by a big mannequin to detect any remaining delicate content material earlier than launch. “This layered course of ensures enterprise-grade robustness for environments dealing with delicate buyer or monetary knowledge,” Wang added.

To speed up analysis, the crew additionally curated AgentNetBench, an offline benchmark that gives a number of appropriate actions for every step, providing a extra environment friendly solution to measure an agent’s efficiency.

A brand new recipe for coaching brokers

The OpenCUA framework introduces a novel pipeline for processing knowledge and coaching computer-use brokers. Step one converts the uncooked human demonstrations into clear state-action pairs appropriate for coaching vision-language fashions (VLMs). Nonetheless, the researchers discovered that merely coaching fashions on these pairs yields restricted efficiency good points, even with massive quantities of information.

The important thing perception was to enhance these trajectories with chain-of-thought (CoT) reasoning. This course of generates an in depth “interior monologue” for every motion, which incorporates planning, reminiscence, and reflection. This structured reasoning is organized into three ranges: a high-level statement of the display screen, reflective ideas that analyze the state of affairs and plan the following steps, and eventually, the concise, executable motion. This method helps the agent develop a deeper understanding of the duties.

“We discover pure language reasoning essential for generalizable computer-use basis fashions, serving to CUAs internalize cognitive capabilities,” the researchers write.

This knowledge synthesis pipeline is a common framework that may be tailored by corporations to coach brokers on their very own distinctive inside instruments. Based on Wang, an enterprise can file demonstrations of its proprietary workflows and use the identical “reflector” and “generator” pipeline to create the required coaching knowledge. “This permits them to bootstrap a high-performing agent tailor-made to their inside instruments while not having to handcraft reasoning traces manually,” he defined.

Placing OpenCUA to the take a look at

The researchers utilized the OpenCUA framework to coach a variety of open supply VLMs, together with variants of Qwen and Kimi-VL, with parameter sizes from 3 billion to 32 billion. The fashions had been evaluated on a collection of on-line and offline benchmarks that take a look at their skill to carry out duties and perceive GUIs.

The 32-billion-parameter mannequin, OpenCUA-32B, established a brand new state-of-the-art success charge amongst open supply fashions on the OSWorld-Verified benchmark. It additionally surpassed OpenAI’s GPT-4o-based CUA and considerably closed the efficiency hole with Anthropic’s main proprietary fashions.

For enterprise builders and product leaders, the analysis provides a number of key findings. The OpenCUA technique is broadly relevant, bettering efficiency on fashions with totally different architectures (each dense and mixture-of-experts) and sizes. The skilled brokers additionally present robust generalization, performing properly throughout a various vary of duties and working techniques.

Based on Wang, the framework is especially suited to automating repetitive, labor-intensive enterprise workflows. “For instance, within the AgentNet dataset, we already seize just a few demonstrations of launching EC2 situations on Amazon AWS and configuring annotation parameters on MTurk,” he advised VentureBeat. “These duties contain many sequential steps however observe repeatable patterns.”

Nonetheless, Wang famous that bridging the hole to dwell deployment requires addressing key challenges round security and reliability. “The most important problem in actual deployment is security and reliability: the agent should keep away from errors that would inadvertently alter system settings or set off dangerous unwanted side effects past the meant job,” he mentioned.

The researchers have launched the code, dataset, and weights for his or her fashions.

As open supply brokers constructed on frameworks like OpenCUA develop into extra succesful, they might basically evolve the connection between information employees and their computer systems. Wang envisions a future the place proficiency in complicated software program turns into much less vital than the power to obviously articulate targets to an AI agent.

He described two main modes of labor: “offline automation, the place the agent leverages its broader software program information to pursue a job end-to-end,” and “on-line collaboration, the place the agent responds in real-time and works facet by facet with the human, very similar to a colleague.” Mainly, the people will present the strategic “what,” whereas more and more refined AI brokers deal with the operational “how.”