Computer Use Agents (CUAs)—AI models that can navigate desktops, browsers, and mobile apps—are one of the most exciting frontiers in automation. However, the field currently suffers from a "fragmentation tax." Every new environment, dataset, or model comes with its own custom code, making it incredibly difficult to scale training or compare results across different platforms.
CUA-Lite is a new open platform designed to solve this by providing a unified infrastructure for developing, evaluating, and training CUAs. By standardizing the "plumbing" of agent development, the authors have created a playground where researchers can easily move between desktop, web, and mobile tasks using a single, consistent interface.
TL;DR
CUA-Lite standardizes the development of computer use agents by providing a unified environment interface, a common data format, and shared model harnesses. It introduces lightweight, VM-free Docker sandboxes that are 6x more memory-efficient than traditional virtual machines while maintaining high evaluation fidelity.
What Problem Does This Paper Address?
Developing CUAs is currently a mess of incompatible resources. If you want to train an agent on a desktop benchmark (like OSWorld) and then test it on a mobile one (like AndroidWorld), you typically have to rewrite the action mapping, context management, and rollout loops from scratch.
Furthermore, the infrastructure for Supervised Fine-Tuning (SFT), Reinforcement Learning (RL), and evaluation is often siloed. This fragmentation prevents researchers from easily scaling their models across different platforms and makes it nearly impossible to share training pipelines. Finally, existing desktop environments often rely on heavy Virtual Machines (VMs) that are slow to boot and expensive to run in the cloud.
Methods
The authors introduce three core abstractions to unify the field:
- Lite.Gym: A unified environment interface that wraps 12 different benchmarks into a standard Gym-style
reset/steploop. It standardizes GUI actions (like click, type, and swipe) and observations (screenshots and text) across all platforms. - VM-Free Sandboxes (Lite.*): To replace heavy VMs, the team reverse-engineered the OSWorld environment into lightweight Docker containers. These "sandboxes" share the host kernel, drastically reducing the resource overhead.
- Model Harnesses: A standardized "scaffolding" for each model family (supporting 14 families including GPT-4/5, Claude, and Qwen). The harness handles the translation between the model's raw output and the standardized
Lite.Gymactions, allowing the same code to be used for evaluation, SFT, and RL.
Datasets
CUA-Lite integrates a massive amount of data into its unified Lite.Sample format:
- 12 Benchmarks: Including OSWorld, WebVoyager, AndroidWorld, and Mind2Web.
- 1.26 Million Converted Samples: Sourced from 9 major public CUA datasets.
- 82,000 Fresh Rollouts: Recorded from frontier models like GPT-5.5 and Qwen on verifiable training tasks.
- 30,000+ Verifiable Tasks: Specifically designed for training, including synthesized and perturbed tasks to prevent overfitting.
Key Insights and Ideas
- Docker is Enough: One of the most practical findings is that you don't need a full VM for high-fidelity desktop agent research. The Docker-based
Lite.OSWorldachieved a 0.98 Spearman correlation with the original VM-based benchmark while using only 1/6th of the memory. - The "Thinking" Boost: For top-tier models like GPT-5.5, enabling reasoning (Chain-of-Thought) leads to significantly higher success rates. Interestingly, these models also produce longer rollouts as they "verify" their work before finishing.
- Unified Training Works: The authors demonstrated a "recipe" for improving open-weight models: first SFT a small model (Qwen3.5-4B) on high-quality rollouts from a "teacher" model (GPT-5.5), then further refine it using RL (specifically the GRPO algorithm) on verifiable tasks.
Key Results
- Performance Gains: SFTing the Qwen3.5-4B model on GPT-5.5 rollouts improved its desktop success rate from 22.6% to 44.8% and its browser success rate from 4.4% to 60.0%.
- RL Improvement: Applying GRPO (Reinforcement Learning) further boosted the SFTed model's performance, raising the browser success rate from 60.0% to 72.8%.
- Reasoning Ablation: For open-weight models, the impact of reasoning is mixed; it sometimes shortens rollouts but doesn't always improve scores as dramatically as it does for frontier API models.
Limitations and Future Work
- Mobile Still Needs Emulators: While desktop was successfully "Dockerized," mobile environments in CUA-Lite still require heavy emulators.
- Linux-Centric: The lightweight sandboxes currently only support Linux; Windows-based environments like WindowsAgentArena still require traditional VMs.
- Verifier Accuracy: The authors note that they haven't fully audited all training verifiers, and some tasks rely on model-based judges which can be imperfect.
- Generalization: The training in this paper focused on the same applications used in evaluation; future work is needed to see how well these gains transfer to entirely unseen software.
Summary updated