Preprint: OS-Helix Unifies GUI and Code Execution, Reaching 66.44% Success in Desktop Automation
New research shows adding code tools naively degrades GUI agents, requiring hybrid training for native coordination. OS-Helix achieves significant breakthroughs on the OSWorld benchmark.
ImportanceMaterialEvidenceE2 unreplicatedWrite-upQuick
OS-Helix achieved a 66.44% success rate on the OSWorld-Verified benchmark, representing a 42.9% relative improvement over matched Qwen3.5 hybrid baselines. These results come from a preprint submitted on October 15, 2026, titled "OS-Helix: Unifying GUI Interaction and Code Execution in Native Computer-Using Agents" by Bowen Yang et al.
Traditional approaches attempt to combine GUI operations and code execution tools through prompt engineering, but the study reveals that such "prompt-level tool composition" cannot achieve effective native coordination. The authors note that simply adding a code tool to a GUI agent actually degrades its performance, indicating that GUI/code routing decisions must be internalized through end-to-end training on high-quality hybrid data.
To address supervision challenges, the team introduced "application-artifact composition," pairing desktop applications with authentic file artifacts to instantiate executable Ubuntu states. They also employed a decoupled instruction/evaluator generation pipeline to ensure task final states could be verified regardless of the solution modality. The model was cold-started with SFT and further improved through online reinforcement learning in live desktop environments.
These results are from a first-party benchmark (E2) and have not yet been independently reproduced. The core contribution lies in demonstrating that simple tool chaining is insufficient for complex desktop automation tasks, necessitating specialized training on a unified action space.