AI Paper Atlas

Research map / Multi-agent systems and games / Human-AI and goal-conditioned

ShowUI-p: Flow-based Generative Models as GUI Dexterous Hands

Siyuan Hu, Kevin Qinghong Lin, Mike Zheng Shou

CVPR 2026 · Multi-agent systems and games › Human-AI and goal-conditioned · approach: Embodied agent learning

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use flow matching action expert instead of diffusion policy (ablation)ShowUI-π (SmolVLM backbone) · ScreenDrag offlineendpoint accuracy (%): 47.33 -> 78.55Improved
use flow matching instead of language-token action modeling (ablation)ShowUI-π (SmolVLM backbone) · ScreenDrag offlineendpoint accuracy (%): 0.40 -> 78.55Improved
add start/end loss reweighting (scale 10) (ablation)ShowUI-π · ScreenDrag online overalltask success rate (%): 10.49 -> 26.98Improved
increase reweighting scale from 10 to 15 (ablation)ShowUI-π · ScreenDrag online overalltask success rate (%): 26.98 -> 20.80Hurt
add directional regularization (lambda 0.1) (ablation)ShowUI-π · ScreenDrag online overalltask success rate (%): 12.63 -> 26.98Improved
increase action chunk size from 10 to 20 (ablation)ShowUI-π · ScreenDrag offline, exec step 1endpoint accuracy (%): 75.45 -> 78.55Improved
execute more steps (5 vs 1) before re-observation (ablation)ShowUI-π chunk 10 · ScreenDrag offlineendpoint accuracy (%): 75.45 -> 66.34Hurt
larger chunk size 20 when executing 5 steps (ablation)ShowUI-π · ScreenDrag offline, exec step 5endpoint accuracy (%): 66.34 -> 79.22; error 241.60 -> 205.23Improved
use unified click/drag head instead of separate heads (ablation)ShowUI-π · ScreenDragonline drag success rate (%): 23.25 -> 26.98 (offline 79.22 -> 78.55)Improved
scale proprietary or open baseline size (OpenCUA-7B vs 32B)OpenCUA · ScreenDrag onlinetask success rate (%): 20.79 (32B) vs 21.98 (7B)No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy