UC Berkeley Researchers Release CUA-Lite, an Open Platform Unifying Sandboxes, Data, Evaluation and RL for Computer-Use Agents
UC Berkeley researchers released CUA-Lite, an open platform that unifies computer-use agents (CUAs) by standardizing agents, environments, traces, and evaluation/training frameworks under one action space and data schema Lite.OSWorld eliminates the "VM tax" by running OSWorld tasks in plain Docker containers instead of QEMU/KVM VMs, reducing memory from 4.1 GB to 0.9 GB and enabling ~4.6× more parallel instances Cross-fidelity validation across 13 models confirms that scores and training signals
Analysis
TL;DR
- UC Berkeley researchers released CUA-Lite, an open platform that unifies computer-use agents (CUAs) by standardizing agents, environments, traces, and evaluation/training frameworks under one action space and data schema
- Lite.OSWorld eliminates the "VM tax" by running OSWorld tasks in plain Docker containers instead of QEMU/KVM VMs, reducing memory from 4.1 GB to 0.9 GB and enabling ~4.6× more parallel instances
- Cross-fidelity validation across 13 models confirms that scores and training signals earned in containers transfer directly to the real OSWorld VM benchmark
- The platform ships LiteSample (a unified supervised-learning schema), per-model adapters, 30k+ verifiable tasks, 15+ benchmarks, 10+ agents, and 20+ datasets published free on Hugging Face
- CUA-Lite supports end-to-end SFT and RL training behind a single command interface, with demonstrated results including Qwen3-VL-2B fine-tuning and GRPO-based reinforcement learning on mobile tasks
Why It Matters
CUA-Lite addresses a critical infrastructure bottleneck in the computer-use agent ecosystem: the fragmentation of agents, environments, traces, and evaluation tools across incompatible repositories. By unifying these components under a single schema and action space, it dramatically lowers the barrier to entry for researchers and practitioners building and benchmarking CUAs. The container-based approach also makes large-scale training and evaluation feasible on standard cloud infrastructure without requiring specialized virtualization support.
Technical Details
- Lite.OSWorld: Replaces full QEMU/KVM virtual machines with lightweight Docker containers running GNOME desktops, achieving identical task suites and evaluators while reducing memory footprint by 78% (4.1 GB → 0.9 GB) and cold start time by 6 seconds. Parallelism increases ~4.6× with no /dev/kvm or nested virtualization required.
- LiteSample Schema: A single supervised-learning data format (plain parquet plus images) shared across all environments, agent types, and task categories. Over ten existing CUA datasets (Aguvis, OpenCUA, ScaleCUA, GUI-360, GUIOdyssey, Multimodal-Mind2Web) have been preprocessed into this schema and published on Hugging Face.
- Per-Model Adapters: The framework includes adapters that pack unified LiteSample data into each model family's native training format, with history collapsing to allow multiple steps within a single forward pass.
- Unified Training Loop:
lite.gymprovides a consistent interface (screenshots in, actions out) with one action space per platform. The same loop supports both supervised fine-tuning (SFT) and reinforcement learning (RL via GRPO on Slime). Demonstrated SFT on Qwen3-VL-2B-Instruct lifted mean episode return from 0.138 to 0.237 on the 332-task lite.osworld eval split. - Breadth of Coverage: The platform includes Lite.ScaleCUA, Lite.CUAGym, and Lite.CUAWorld (expanding to ~40 applications including Blender, QGIS, and VS Code), 15+ benchmarks spanning grounding, desktop, browser, and mobile domains, and 10+ built-in agents (GPT, Claude, Gemini, Qwen3-VL, UI-TARS, Fara-7B, MAI-UI, and others).
Industry Insight
- The containerization of OSWorld-style benchmarks represents a significant scalability breakthrough, enabling researchers and companies to run large-scale CUA training and evaluation on commodity cloud infrastructure without specialized virtualization hardware—potentially accelerating the pace of agent development across the industry.
- The unified LiteSample schema and per-model adapter pattern could become a de facto standard for CUA data interoperability, reducing fragmentation and enabling cross-dataset training that was previously impractical due to incompatible formats.
- While the platform shows promising results, the SFT demonstration was based on a single configuration on two GPUs without independent reproduction, and the repository currently lacks an explicit license—practitioners should verify licensing terms before commercial deployment and treat early performance claims as preliminary until further validation.
Disclaimer: The above content is generated by AI and is for reference only.