WinDOM: Self-Family Distillation for Small-Model GUI Grounding

A new corpus and training method aim to improve small-scale GUI-grounding agents without expensive human annotation, according to a paper posted to arXiv on June 24, 2026 [1]. The work introduces WinDOM, a dataset of 54,425 records harvested from a headless Windows 11 web reimplementation, and Self-Family Distillation, a technique for combining supervised fine-tuning with reinforcement learning [1]. The paper, authored by Chengheng Li Chen, targets models with roughly 2 billion parameters, a scale the authors describe as attractive for on-device deployment and accessibility tooling [1]. At this size, two open questions persist: how to obtain bounding-box training data without costly human annotation, and how to effectively combine supervised fine-tuning with reinforcement learning [1]. WinDOM addresses the data question by driving an open-source Windows 11 web reimplementation under headless Playwright, reading bounding boxes directly from the Document Object Model without optical character recognition or human labeling [1]. The resulting corpus contains 54,425 records [1]. For training, the paper proposes Self-Family Distillation, or SFD, a single rejection-sampling cold-start parameterized only by the choice of teacher [1]. The teacher can be either an exponential moving average of the student model, requiring no external model, or a frozen larger model from the same family [1]. The researchers then treat the saturation depth of the SFD cold-start as an explicit hyperparameter for Group Relative Policy Optimization, or GRPO [1]. On a Qwen3.5-2B student, the under-saturated cold-start proved a better GRPO initializer than the fully converged one [1]. The SFD-4B variant with early-init reinforcement learning posted a gain of 5.4 points in out-of-distribution mean over the base model, with component gains of 3.5 on ScreenSpot-Pro, 7.0 on OSWorld-G, and 5.8 on ScreenSpot-V2 [1]. The same-size EMA mode reached an OOD-mean of 65.2, within roughly one point of the cross-size 4B variant at 66.3, without relying on an external teacher [1]. The paper appeared on arXiv, the open-access e-print repository that has hosted scientific preprints since 1991 and now receives approximately 24,000 submissions per month [6]. The repository is not peer-reviewed but serves as a primary distribution channel in fields such as computer science and physics [6]. The work was submitted on June 24, 2026, in the Artificial Intelligence category [1].