The preprint reports a lower rate of annotated cross-step control failures for LocalLSTC than for the average of four baseline systems. At least one such failure appeared in 43.1% of LocalLSTC trajectories on OSWorld, compared with 59.4% for the four-baseline average. The reported gap was 16.3 percentage points, with a bootstrap 95% confidence interval of 8.3 to 24.2 points. The NONE label appeared in 42.3% of LocalLSTC trajectories.
The system's central design separates persistent Long-Term Control from bounded Short-Term Execution. Long-to-Short Planning forms the current commitment, while Short-to-Long Control uses execution outcomes to update persistent state.
Results varied by benchmark and backbone
The evaluation covered 369 Linux tasks in OSWorld and 154 Windows tasks in WindowsAgentArena. The default configuration used Qwen3.5-9B as its planner and GTA1-7B as its visual grounder. OSWorld configurations were run independently twice, while other reported results used one run.
Using Qwen3.6-27B with GTA1-7B, LocalLSTC reported OSWorld success rates of 64.7% at SR-100, 62.7% at SR-50 and 51.4% at SR-15. On WindowsAgentArena, the corresponding rates were 65.3%, 63.4% and 48.5%. The WindowsAgentArena SR-50 result was reported as 18.1 percentage points above the previous local result and within 0.1 points of the strongest API result.
The reported averages also varied with the reasoning backbone. Across four frameworks, average OSWorld SR-100 was 60.9% with GPT-5 and 37.7% with Qwen3.5-9B, while framework-level differences ranged from 21.1 to 25.0 percentage points.
What the annotations found
Among failed baseline trajectories, at least one annotated control failure appeared in 91.6%. RECOVER, STOP and LOOP were the most frequent labels, occurring in 64.2%, 41.1% and 34.4% of those trajectories, respectively.
The labels were checked on a stratified sample of 37 OSWorld tasks, yielding 185 trajectories. Two PhD-level researchers independently annotated the sample after method identity was removed. They agreed on the presence or absence of an ANY label 86.5% of the time, with Cohen's kappa of 0.730. GPT-5.6-sol's ANY label had F1 scores of 86.1% and 85.2% against the two researchers, with kappa values of 0.686 and 0.647.
For the failure comparison, the systems were evaluated on the same tasks, allowing paired binary outcomes. The reported 95% confidence intervals used a task-level percentile bootstrap with 10,000 resamples and random seed 2027.
Configuration comparisons produced different scores
In grouped ablations, the full LocalLSTC configuration scored 49.1% on SR-100. The version without Long-to-Short Planning scored 36.4%, while the version without Short-to-Long Control scored 41.5%. These are differences between tested configurations and do not, on their own, establish that either omission caused the gap.
Separate component comparisons reported SR differences of 17.5 percentage points when SA was omitted, 11.1 points for the persistent subgoal, 8.8 for Final Verification, 8.4 for state-conditioned routing, 8.4 for candidate proposals, 5.6 for the multi-action list and 4.0 for stall/loop handling. These contrasts describe tested versions rather than isolating the causal contribution of any single component.
Recovery and runtime events
Final Verification rejected termination on 18.3% of OSWorld tasks. Among the reported OSWorld recovery cases, 63.8% subsequently succeeded after returning to execution. On WindowsAgentArena, 21.4% of tasks entered recovery; among the reported recovery cases, 51.5% subsequently succeeded after an average of 5.2 additional steps.
The grouped runtime comparison found similar stall or repetition incidence: 4.5% for the full configuration, 4.3% without Long-to-Short Planning and 4.6% without Short-to-Long Control. Conditional evaluator scores after a stall were 34.3%, 18.8% and 11.8%, respectively. The full configuration's post-recovery score was 34.0%, compared with 20.9% for the version without Long-to-Short Planning; its rejected-termination score was 38.9%, compared with 23.1% for the same version.
Evidence remains benchmark-bound
The manuscript treats temporal organization of control information as a distinct design dimension for locally deployed GUI agents. The evidence is benchmark-based: 369 Linux tasks in OSWorld and 154 Windows tasks in WindowsAgentArena, run with selected model configurations. It describes performance in those settings, not outcomes for human users or real-world deployment.
One practical caution is that the run schedule was uneven: OSWorld configurations were run twice independently, while other results used one run. The failure-label validation covered 37 tasks and 185 trajectories, and agreement was good but not perfect. Those details leave open whether the pattern would hold across more systems, environments and longer tasks.
The manuscript is labeled arXiv:2608.25777v1 and dated 26 August 2026. The authors state that they have released the implementation, task-level outputs and evaluation artifacts through an anonymous code repository, with appendices documenting the annotations, implementation and prompts.
Paper data and sources
Original title: LocalLSTC: A Long Short-Term Control Architecture for Locally Deployed GUI Agents
Authors: Weiming Li, Helen Paik, Yulei Sui
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text