OpenGCUA

TL;DR

OpenGCUA combines visual interaction and code execution, letting the model choose more of its observation and task-state decisions. The authors report improved OSWorld-v2 partial scores across all tested models. (source)

Authors & Org

Hanzhang Zhou, Long Li, Quyu Kong and colleagues; the captured abstract page lists the complete author roster but does not establish affiliations. Submitted 2026-10-15. (source)

Method

Code execution lets intermediate results determine later operations. Harness-to-model decision transfer moves selected decisions out of fixed scaffolding. The release includes implementation, prompts, evaluation trajectories and preliminary real-device and cross-device tooling. (source)

Results

• Across eight models, the full configuration raises mean partial score by +19.6 points and reduces reported steps per task by 53.0% relative to batched GUI-only execution. Partial score is not a task-completion rate. (source)
• On four real macOS tasks, selected Qwen3.8-Max runs use 19.4× the reported reasoning tokens of GPT-6 Astra. The abstract does not establish a general cost ratio. (source)

Significance

For Agents (LLM Agents) and Eval Harness Configuration, the result makes interface choice and decision ownership explicit parts of an evaluation. It does not isolate a model-only capability gain. (source)

Open Questions

Full paper and code were not reviewed. The abstract leaves per-model completion rates, total runtime, tool costs and real-device safety failure rates unspecified. (source)

Cite

arXiv:2610.22200, first submitted 2026-10-15. (source)