Saved in:
Bibliographic Details
Main Authors: Sumyk, Marta, Kosovan, Oleksandr
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2603.10577
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917334624501760
author Sumyk, Marta
Kosovan, Oleksandr
author_facet Sumyk, Marta
Kosovan, Oleksandr
contents Computer-Use Agents (CUAs) are emerging as a new paradigm in human-computer interaction, enabling autonomous execution of tasks in desktop environment by perceiving high-level natural-language instructions. As such agents become increasingly capable and are deployed across diverse desktop environments, evaluating their behavior in a scalable and reliable manner becomes a critical challenge. Existing evaluation pipelines rely on static benchmarks, rule-based success checks, or manual inspection, which are brittle, costly, and poorly aligned with real-world usage. In this work, we study Vision-Language Models (VLMs) as autonomous auditors for assessing CUA task completion directly from observable interactions and conduct a large-scale meta-evaluation of five VLMs that judge task success given a natural-language instruction and the final environment state. Our evaluation spans three widely used CUA benchmarks across macOS, Windows, and Linux environments and analyzes auditor behavior along three complementary dimensions: accuracy, calibration of confidence estimates, and inter-model agreement. We find that while state-of-the-art VLMs achieve strong accuracy and calibration, all auditors exhibit notable performance degradation in more complex or heterogeneous environments, and even high-performing models show significant disagreement in their judgments. These results expose fundamental limitations of current model-based auditing approaches and highlight the need to explicitly account for evaluator reliability, uncertainty, and variance when deploying autonomous CUAs in real-world settings.
format Preprint
id arxiv_https___arxiv_org_abs_2603_10577
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle CUAAudit: Meta-Evaluation of Vision-Language Models as Auditors of Autonomous Computer-Use Agents
Sumyk, Marta
Kosovan, Oleksandr
Artificial Intelligence
Human-Computer Interaction
Computer-Use Agents (CUAs) are emerging as a new paradigm in human-computer interaction, enabling autonomous execution of tasks in desktop environment by perceiving high-level natural-language instructions. As such agents become increasingly capable and are deployed across diverse desktop environments, evaluating their behavior in a scalable and reliable manner becomes a critical challenge. Existing evaluation pipelines rely on static benchmarks, rule-based success checks, or manual inspection, which are brittle, costly, and poorly aligned with real-world usage. In this work, we study Vision-Language Models (VLMs) as autonomous auditors for assessing CUA task completion directly from observable interactions and conduct a large-scale meta-evaluation of five VLMs that judge task success given a natural-language instruction and the final environment state. Our evaluation spans three widely used CUA benchmarks across macOS, Windows, and Linux environments and analyzes auditor behavior along three complementary dimensions: accuracy, calibration of confidence estimates, and inter-model agreement. We find that while state-of-the-art VLMs achieve strong accuracy and calibration, all auditors exhibit notable performance degradation in more complex or heterogeneous environments, and even high-performing models show significant disagreement in their judgments. These results expose fundamental limitations of current model-based auditing approaches and highlight the need to explicitly account for evaluator reliability, uncertainty, and variance when deploying autonomous CUAs in real-world settings.
title CUAAudit: Meta-Evaluation of Vision-Language Models as Auditors of Autonomous Computer-Use Agents
topic Artificial Intelligence
Human-Computer Interaction
url https://arxiv.org/abs/2603.10577