Study Questions Whether More Compute Time Helps Local Computer-Use Agents
A new preprint tests whether extra inference-time computation actually helps small, locally run computer-use agents.
What the researchers studied
A new preprint, posted to arXiv by authors Woongkyu Lee and Jungwook Choi, examines a specific question: deploying autonomous computer-use agents locally matters for privacy, cost, and practical usability, but does giving these resource-constrained local models more computation during a task actually make them better? Prior work had shown that inference-time scaling can help larger, frontier computer-use agents, but the paper says its effectiveness for smaller local models remained poorly understood before this study.
To find out, the researchers ran a systematic empirical study of inference-time scaling in local computer-use agents across four dimensions: contextual, temporal, structural, and parallel. They evaluated three model families on this basis: Qwen3-VL-8B/30B-A3B, UI-TARS-1.5-7B, and OpenCUA-7B, all tested on the OSWorld benchmark.
What they found
The paper reports that additional computation often yields diminishing returns while changing failure modes rather than eliminating them. According to the researchers, contextual scaling, meaning giving the agent more historical grounding from earlier in a task, does improve trajectory stability and task accuracy, but those gains saturate as token cost increases. As that happens, the paper reports, failures shift away from repetitive or stalled trajectories and toward premature false successes, where the agent wrongly reports a task as complete.
Temporal scaling, or letting an agent run for more steps, similarly reduces the agents hitting a maximum-step stall, the researchers write, but does not substantially improve whether the task actually succeeds. The paper interprets this as evidence that longer horizons often extend an already-wrong trajectory rather than correcting it.
The researchers further report that structural decomposition, splitting a task into planning and execution stages, can introduce its own planning and formatting overhead in local two-stage agents. Parallel scaling, running multiple attempts at once, partially mitigates these failures, according to the paper, but only at a substantial computational cost.
What the researchers recommend
Based on these results, the paper concludes that efficient local computer-use agents require selective compute allocation, failure-aware control mechanisms, and agentic frameworks designed specifically around the capabilities and limitations of local models.
Key facts
- The study evaluates Qwen3-VL-8B/30B-A3B, UI-TARS-1.5-7B, and OpenCUA-7B on the OSWorld benchmark.
- The researchers tested four scaling dimensions: contextual, temporal, structural, and parallel.
- Contextual scaling's gains saturate and failures shift toward premature false successes, the paper reports.
- Temporal scaling reduces stalls but does not substantially improve task success, according to the researchers.
- Parallel scaling can offset some failures but at a substantial computational cost, the paper says.
Check your understanding
Which four dimensions of inference-time scaling did the researchers test in local computer-use agents? According to the paper, what kind of failure increases as contextual scaling's gains saturate?