$ cat wiki/papers/2026/2610.41432-hybrid-computer-use.md
Hybrid Computer Use
TL;DR
HCU combines code, accessibility and graphical interaction so agents can choose channels rather than repeatedly reconstructing everything from screenshots. The authors report better task success and efficiency across the models they evaluate. (source)
Authors & Org
Haoqing Wang, Xiaoyi Zhang, Sicheng Fan, Zhexin Zhang, Xinjie Zhang, Jiahao Li, Bin Li and Yan Lu. Affiliations are not established by the captured metadata. First submitted 2026-10-30. (source)
Method
The interface combines programmatic code and command-line tools, semantic accessibility (AX), and visual graphical user interfaces (GUI). AX tools provide compact observations, batched actions and differences describing their effects. A companion skill coordinates channel selection, with guards against acting on stale state when switching channels. (source)
Results
On OSWorld 2.0, HCU has the highest Partial and Binary scores among the compared systems, according to the authors. All evaluated models improve but distribute work differently across channels. The abstract gives no numerical scores or latency measurements. (source)
Significance
For Agents (LLM Agents) and Eval Harness Configuration, available interaction channels are part of the evaluated system. This motivates reporting the interface alongside the model; it does not establish universal superiority across applications. (source)
Open Questions
The captured abstract does not quantify the contribution of each channel or the overhead of stale-state guards. Full-paper ablations and code were not reviewed. (source)