Alibaba's adaptive GUI agent training cuts environment interactions by 56-61%
Yuxiang Chai · hf · 2026-11-04
Alibaba Token Hub and CUHK MMLab propose a two-stage adaptive training framework for multi-step GUI agents, addressing where to learn and how far to roll out. Trajectory-level GRPO is inefficient: task-level rewards give coarse supervision over intermediate decisions while mastered prefixes are repeatedly re-executed.
- Stage I: build reusable intermediate checkpoints from successful expert trajectories, select informative pivots via action disagreement (uncertainty measured directly from sampled GUI actions), and perform one-step RL at these pivots.
- Stage II: progressively expand policy responsibility via adaptive Local, Suffix, and Root training actions tied to task progress.
On Mobile-Gym and CUA-Gym-Hub, the method reduces environment interactions by 56-61% versus vanilla GRPO while improving task performance.
More from coding & agent
- Axion 1.0: ByteDance-Backed Open Agent Foundation Model With 3-Stage Recipe — Yuqi Jia · 2026-11-04
- Caper Pre-Executes Shell Commands in Sandboxes for Safer Agent Approvals — SingaporeManagementUniversity · 2026-11-04
- Claude Code lead: AI shifts team bottlenecks upstream to PMs and downstream to verification — Flomerboy · 2026-11-04
- 10 open-source repos that run AI agents while you sleep: cron, webhooks, cloud PRs — victor_explore · 2026-11-04
- The shopping cart pattern: recursively chaining AI micro-decisions to handle multi-intent queries — hardimanjames · 2026-11-04
- Researcher: Codex is 'miles ahead' of Claude Code on computer use — Yuchenj_UW · 2026-11-04