纯 GUI 的操作方式，对越来越强的模型反而是束缚。
This is a AI post classified by AGK Intelligence as AI agents (research), kept by the AI Radar because it carries real work, not commentary.
纯 GUI 的操作方式，对越来越强的模型反而是束缚。 MSRA 这份 Mage-VL-Cowork 做的就是 hybrid computer use：把一个本来没有 agentic CUA 能力的 4B 小模型，用五阶段课程从 midtraining 一路走到 SFT 和 RL，最后在 OSWorld-Verified 上拿到 66.0%。 最重的活是数据。Mage-CUA-Forge 用可复用的、按技能条件化的生成流程替代了逐任务造 agent，覆盖 6,915 个技能、合成 293,806 个桌面和 Web 任务，同时把单个任务的重复开销压低了 82.6%。 另外两个选择值得记：一是不直接蒸馏一个默认 teacher，而是把 teacher 的 harness 进化出来，再实例化专门的 CUA harness，用它们产出高质量轨迹和分阶段监督；二是 harness 本身按小模型重新设计，探索 CLI 和 GUI 混合执行，在复杂能力和执行效率之间找平衡。 结果是一张挺好看的图：4B 站到了 OSWorld-Verified 66.0 的帕累托前沿上。 值得说的是这个组合本身。4B 能上前沿，靠的不是参数量，是环境、harness 和分阶段训练被完整搭了起来。纯 GUI 那套 human bias 在弱模型上是唯一选择，在强模型上就成了成本。 https://arxi
Posted by 青稞社区 (3.7k followers) 17 h ago · 11 likes · 690 views · view the original post on X. Kept by the AI Radar as AI agents. Tools mentioned: arXiv.org.
More AI work like this
- If you want to learn how to build self-improving harnesses, check out this paper. — @dair_ai
- I asked what the oldest or weirdest device talking to your Hermes Agent is. my favorites: — @witcheer
- Have two agents doing a thorough code review of the new Hermes Agent Templates to ensure… — @MichaelGannotti
- Your AI agent needs to browse the web. Does it need a full Chromium instance to do it? — @mr_r0b0t
- Copy a WordPress contact page from an old site and its SiteOrigin form still sends to… — @johnturner
- Tobira turns your website into an agent that sells to your customer's AI inside ChatGPT… — @ProductHunt
- how we’re building decision models at @liquidai — @helloiamleonie
- We’ve been hard at work trying to make models work with harnesses better. The best thing… — @LysandreJik
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 65.6k posts from 5k X accounts over the last 14 days, 10.4k tools, 19 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-06 16:05 UTC. Full method.