UC Berkeley Researchers Release CUA-Lite, an Open Platform Unifying Sandboxes, Data, Evaluation and RL for Computer-Use Agents
2026-09-06 14:11Models🔥 42.2 heat score
1sources
1days unfolding
42.2heat score
4mentions
SummaryAI generated
On September 6, 2026, a team from the University of California, Berkeley, released the CUA-Lite open platform, aimed at unifying the training, evaluation, and data infrastructure for Computer-Use Agents. This platform migrates Ubuntu desktop tasks that originally relied on QEMU/KVM to run in Docker containers through the Lite.OSWorld component, significantly reducing resource consumption and enhancing parallel capabilities. Specifically, memory usage was reduced from 4.1 GB to 0.9 GB, parallel processing capacity increased by approximately 4.6 times, while maintaining scoring consistency. CUA-Lite includes sandbox components such as Lite.ScaleCUA, Lite.CUAGym, and Lite.CUAWorld, covering over 30,000 verifiable tasks and more than ten preprocessing datasets. Additionally, the platform provides a unified data sample format, LiteSample, and an adaptation layer, supporting formats including GPT, Cla…
The team at UC Berkeley released CUA-Lite, a unified open platform for agents used in computing, featuring unified sandboxes, data, evaluation, and reinforcement learning. This platform addresses the fragmentation of existing tools through a single action space and data architecture. Its core component, Lite.OSWorld, migrates Ubuntu desktop tasks that originally required QEMU/KVM to Docker containers, reducing memory usage from 4.1 GB to 0.9 GB while increasing parallelism by approximately 4.6 times and maintaining consistent scoring. The platform includes sandboxed components such as Lite.ScaleCUA, Lite.CUAGym, and Lite.CUAWorld, covering over 30,000 verifiable tasks and more than ten pre-processed public datasets. Additionally, CUA-Lite provides a unified data sample format, LiteSample, and an adaptation layer, supporting training of over ten models such as GPT and Claude within the lite.gym framework…