A computer-use agent can report success after a plausible click even when the application never committed the intended change. VeriCUA treats verification as a second task: inspect interface and structured state, gather receipts and abstain when evidence is insufficient. Its benchmark accuracy improves on several evaluated environments, but its decided…
Inspect the ClaimsCase · DiggingBeagle record
VeriCUA verifies computer-using agents by exploring state instead of trusting screenshots
A verification system checks structured application state and selectively abstains, improving correctness on tested GUI-agent benchmarks at added cost.
Scope
Accuracy, abstention and cost are benchmark-specific; the study does not demonstrate production loss prevention.
Assessment profile
Assessment method- Setting
- evaluation
- Exploitation
- demonstrated
- Evidence
- Grade C
- Remediation
- not applicable
- AI role
- WITH AI
Basis and provenance
The pinned Claim and its primary disclosure support the described controlled result, but do not establish a population-wide rate or calibrated impact. The source is primary to the described analysis, not independent replication of every outcome. Impact remains unbanded because no calibrated impact rubric for this particular evidence is established.
Assessed Nov 2, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-2026-NOV02-VERICUA-STATE-VERIFICATION-A · VeriCUA: Verification as Agentic Exploration for Computer-Using Agents
At a glance
Independent outcome verification is more reliable than trusting agent reports or screenshots alone, but abstention and measurement cost must be reported.
Read the implicationsAccuracy, abstention and cost are benchmark-specific; the study does not demonstrate production loss prevention. The results depend on benchmark instrumentation and structured state access; no customer transaction or safety-critical application validation is provided.
Limits and uncertaintyFull account
A computer-use agent can report success after a plausible click even when the application never committed the intended change. VeriCUA treats verification as a second task: inspect interface and structured state, gather receipts and abstain when evidence is insufficient. Its benchmark accuracy improves on several evaluated environments, but its decided subset and added latency differ from the always-deciding baselines. It does not guarantee that every proprietary application exposes trustworthy state.
VeriCUA's AndroidWorld online result is 90.5% accuracy overall versus 78.4% for its VLM-judge baseline, while the system makes explicit decisions on 82% of that sample. The decided subset reaches 95.1% accuracy in the authors' table, which is not an always-deciding score and must not be compared as such. The extra exploration also costs more: the cited online comparison uses roughly 39.2K tokens and 97 seconds versus 16.2K and 60 seconds for a generated verifier. The outcome is a trade-off between stronger evidence, abstention and cost, not proof of universal GUI completion.
Mechanism and trust boundary
- 01
Agent performs GUI actions
Clicks and text input produce a visible but potentially misleading page state.
Boundary: User intent into application interaction
- 02
Verifier gathers independent evidence
A separate process inspects accessible structured state or receipts.
Boundary: Visual appearance into durable application facts
- 03
Decide or abstain
The verifier labels completion only when its evidence supports it.
Boundary: Observed state into user-facing success claim
Claims & evidence
CLM-2026-NOV02-VERICUA-STATE-VERIFICATION-AThe paper reports 90.5 percent AndroidWorld verification accuracy for VeriCUA versus 78.4 percent for a vision-language judge, with VeriCUA deciding on 82 percent of that sample.supported
Basis: reported finding
- supportsVeriCUA: Verification as Agentic Exploration for Computer-Using Agentsprimary disclosure
Method: Evidence Discipline; Online Verification and Experiments, Table 1
CLM-2026-NOV02-VERICUA-STATE-VERIFICATION-BIn the authors' AndroidWorld online evaluation, VeriCUA reached 90.5 percent overall accuracy with an 82 percent decision rate, versus 78.4 percent accuracy with 100 percent decisions for the VLM judge.supported
Basis: reported finding
- supportsVeriCUA: Verification as Agentic Exploration for Computer-Using Agentsprimary disclosure
Online Verification, Table 1, AndroidWorld accuracy and decision rates
Implications
Independent outcome verification is more reliable than trusting agent reports or screenshots alone, but abstention and measurement cost must be reported.
Controls and mitigations
- Use app-level receipts, checked postconditions and explicit unknown status when protected state cannot be independently inspected.
- Publish verification coverage, abstentions, false positives and cost together; verify a state-changing action against an application receipt rather than a screenshot alone.
Unknowns and contradictions
- Accuracy, abstention and cost are benchmark-specific; the study does not demonstrate production loss prevention.
- The results depend on benchmark instrumentation and structured state access; no customer transaction or safety-critical application validation is provided.
Sources and citation
Material revision history
- Nov 2, 2026 · Canonical change recorded · new in release · revision 214