How Benchmarks Mis-Score Computer-Use Agents

Source

Source name: arXiv
Source kind: paper
Authors / organization: Zihan Dong, Zhiyuan Ma, Zekun Wang, Yunqing Li, Zirou Liu, Ruixuan Deng, Qishi Zhan, Rui Qian
Venue: arXiv: cs.AI
Published: 2026-07-30
Collected: 2026-07-31
URL: http://arxiv.org/abs/2607.28367v1
PDF: https://arxiv.org/pdf/2607.28367v1
DOI: UNKNOWN

Plain-Language Brief

The paper discusses issues with the scoring of Computer-Use Agents (CUAs) using brittle scripted oracles, highlighting problems such as outdated tasks, omitted decisive visual evidence, and hidden failure causes in aggregate reports. The study audits 150 public failure-scored trajectories from various benchmarks and finds that around 15.3% of FAIL verdicts are incorrect, suggesting the need for a more reliable framework for CUA evaluation.

What the Document Says

Computer-use agents (CUAs) are being used to browse the web and operate desktop software.
Their scores are often produced by brittle scripted oracles.
The paper identifies problems in scoring, including outdated tasks, omitted decisive visual evidence, evaluator errors, and hidden failure causes.
A reliability framework is proposed covering task construction, trajectory observation, scoring, and reporting.
150 public failure-scored trajectories from five web, enterprise-workflow, and desktop-control benchmarks were audited.
15.3% of FAIL verdicts were found to be incorrect: 10.7% were evaluator false negatives and 4.7% were broken tasks.
For genuine failures, a three-tier diagnostic taxonomy was developed showing that verification/feedback and planning failures dominate execution/grounding errors.
The findings are connected to newer long-horizon CUA benchmarks, and design rules for CUA evaluation are derived.

Key Concepts Mentioned

Computer-Use Agents (CUAs)
brittle scripted oracles
scoring issues
reliability framework
audit
failure-scored trajectories
web benchmarks
enterprise-workflow benchmarks
desktop-control benchmarks
long-horizon CUA benchmarks
design rules

Claims Made by the Document

15.3% of FAIL verdicts are incorrect.
Verification/feedback and planning failures dominate execution/grounding errors.

Evidence Used

Audit of 150 public failure-scored trajectories.
Three-tier diagnostic taxonomy.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Computer-use agents (CUA) are being deployed to browse the web and operate desktop software, yet their benchmark scores are still commonly produced by brittle scripted oracles. A score is the output of a pipeline in which tasks can be stale, trajectories can omit decisive visual evidence, evaluators can reject valid alternatives, and aggregate reports can hide the cause of failure. We organize these problems into a reliability framework spanning task construction, trajectory observation, scoring, and reporting. We then audit 150 public failure-scored trajectories from five web, enterprise-workflow, and desktop-control benchmarks, find that 15.3% of FAIL verdicts are wrong: 10.7% are evaluator false negatives and 4.7% are broken tasks. For genuine failures, a three-tier diagnostic taxonomy shows that verification/feedback and planning failures dominate execution/grounding errors, while a single scalar success rate can not explain. We connect these findings to newer long-horizon CUA benchmarks and derive stage-specific design rules for CUA evaluation.