Berkeley's CUA-Lite: Open Platform Unifies Eval, SFT & RL for Computer-Use Agents
dawnsongtweets · x · 2026-09-01
Dawn Song's team at Berkeley RDI released CUA-Lite, an open platform for computer-use agents (CUAs) that tackles fragmentation across agents, environments, traces, and training frameworks.
Key pieces:
- One standardized interface and action space between agents and environments, spanning desktop, browser, and mobile
- A unified trace format (LiteSample) — one agent's rollouts can train any other agent via lightweight model adapters
- Lite.Gym connects eval → SFT → RL in one framework, with RL methods like GRPO/GSPO built on Slime
- Optional VM-free Docker sandboxes needing no /dev/kvm, using under ¼ the memory/CPU, with many parallel instances per machine
Already integrated: 10+ CUAs (GPT, Claude, Gemini, Qwen, UI-TARS), 15+ benchmarks (OSWorld, WebArena, AndroidWorld), 30K+ verifiable tasks, and 10+ datasets free on Hugging Face, plus a leaderboard with independently reproduced scores. The project is led by PhD student Zhanhui Zhou and is community-driven open source.
Related event: Berkeley Releases CUA-Lite Open Platform for Computer-Use Agents(2 posts)→
More from coding & agent
- Fintech HR built a sourcing agent on 60,000 past applicants — 30 qualified candidates in a month — G0NZ0BOY · 2026-10-16
- Orchestrating a 100-image portrait of humanity for aliens with ChatGPT, Argon and p5.brush — dx8xb · 2026-10-16
- 3x DGX Spark cluster runs GLM-5.3-Flash locally behind a proxy, 1.75M tokens/day at 119W — Teknium · 2026-10-16
- Builder's one rule for trusted desktop agents: read freely, but sending, deleting, buying always ask first — No_Fan5310 · 2026-10-16
- GPT-built unifi-simulator makes MikroTik switches appear inside UniFi Network — Pytlicek · 2026-10-16
- Stanford drops 45-min SpaceX lecture on how Grok Bot's always-on agent harness works — Hesamation · 2026-10-16