Xie and colleagues built ScreenRef from 586K screenshots, yielding 633K referring-expression instances with box or point annotations.
Open dataset and distillation-to-reinforcement pipeline sharpen GUI grounding in 8B model
ScreenRef's 586K screenshots plus a two-stage D2RL training recipe let a Qwen3-VL-based model beat open-data rivals and match closed-source competitors.
Academic
Qingsong Xie · Jiahao Tian · Jiacheng Tang · Zhifeng Li · Madalina Ciobanu · Navan Preet Singh · +2 more
Prentis AI · Westlake University · Titan Holdings
Research Digest··2 min read
The authors introduce ScreenRef, a large-scale GUI grounding dataset spanning desktop, web and mobile, and D2RL, a two-stage post-training method combining on-policy self-distillation with reinforcement learning.
Why this paper
From Prentis AI and 2 others
In one line
A large GUI grounding dataset and a distillation-to-reinforcement learning training method yield state-of-the-art 8B-scale grounding models.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ✓Compute or model size stated (params 8B)
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks (5 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§