AI

GroundGUI: GUI Grounding by High-Quality Data and Distillation-to-Reinforcement Learning

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 16/10/2026, 02:33:24 UTC.

Briefings

GroundGUI-8B, a new GUI grounding model, was trained using the ScreenRef dataset containing 586K screenshots and 633K referring expressions. It achieved 95.2% accuracy on ScreenSpot-V2 and 63.9% on ScreenSpot-Pro, outperforming many open-source models while matching closed-source performance. The model uses Distillation-to-Reinforcement Learning to improve generalization beyond supervised fine-tuning. Code and data are available via GitHub.

Key points

• GroundGUI-8B achieved 95.2% accuracy on ScreenSpot-V2 benchmark using the ScreenRef dataset.
• The model uses Distillation-to-Reinforcement Learning to improve generalization beyond supervised fine-tuning.

Context

GroundGUI-8B represents a significant advancement in GUI grounding by addressing the scarcity of high-fidelity datasets for diverse environments. The model's performance on benchmarks like ScreenSpot-V2 and OSWorld-G demonstrates its capability to handle complex visual tasks across desktop, web, and mobile domains.

Limitations

Performance metrics are limited to the specific benchmarks tested (ScreenSpot-V2, ScreenSpot-Pro, OSWorld-G, MMBench-GUI L2, UI-Vision). The dataset coverage spans desktop, web, and mobile but lacks data on specific device types or operating systems beyond the mentioned domains. Results are based on a single model version and may not reflect real-world deployment conditions.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 25/10/2026, 17:24:59 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeicdgcow33cfamggzrootumt5jub42ix6i5pkqcwazlv3ktkpeargm