New dataset and distillation method boost small GUI-grounding AI models

Researchers have developed WinDOM, a new dataset and method for training small, approximately 2B parameter GUI-grounding agents. The WinDOM corpus, containing over 54,000 records, was generated by automating interactions with a Windows 11 web reimplementation, extracting bounding boxes directly from the Document Object Model without human annotation or OCR. This approach is paired with Self-Family Distillation (SFD), a technique that uses either a student model's own evolving state or a larger, same-family teacher model for training. Experiments show that a Qwen3.5-2B model fine-tuned with SFD-4B and Early-init RL achieved significant gains on various benchmarks, outperforming the base model. AI

IMPACT This research offers a novel approach to training smaller, more efficient AI models for GUI grounding, potentially enabling wider on-device deployment and accessibility tools.

RANK_REASON The cluster contains a research paper detailing a new dataset and training methodology for AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Chengheng Li Chen
• Document Object Model
• GRPO
• Hugging Face
• Playwright
• Qwen3.5-2B
• Self-Family Distillation
• SFD-4B
• WinDOM
• Windows 11

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →