OpenComputer: Verifiable Software Worlds Expose AI Agents' Execution Hallucinations

Forum topic · 小凯 · 2026-05-25

Summary

Computer-use AI agents often claim tasks are complete when they are not — booking the wrong dates, failing to process payments, or skipping steps entirely. This 'execution hallucination' persists because current evaluations rely on LLM judges that assess only screenshots, guessing outcomes instead of checking real system state. The OpenComputer framework (May 2026) tackles this by embedding verifiers directly into software backends. It ships a real-state verifier that compares actual application states against task goals, a self-evolving layer that learns from execution feedback, and a benchmark of 33 desktop applications with 1,000 auditable tasks. Rewards are computed from genuine state checks rather than semantic judgments. Under this stricter evaluation, top models that previously scored near-perfect on older benchmarks saw scores cut roughly in half, revealing that much of their prior success was superficial. This post explains the evaluation bias problem, OpenComputer's ground-truth architecture, and why state-based verification matters for the future of computer-use agents. Reference: arXiv:2605.19769.

When AI assistants claim 'done' but nothing happened

Today's AI agents can control a browser or desktop: clicking, typing, filling forms. Yet sometimes an agent reports "all done!" while the hotel was booked for the wrong date — or no payment was ever processed. This is execution hallucination: the agent believes it finished, and even its LLM-based evaluator agrees, but the real world says otherwise.

The root cause: armchair examiners

Most current computer-use agent benchmarks use another large language model as the judge. These judges are knowledgeable but live outside the software — they cannot see the real database state. If the agent claims it sent an email, the judge looks at a screenshot, nods, and awards full marks.

> Evaluation bias: the referee is guessing, not checking, so scores are inflated.

OpenComputer: ground truth, direct connection

Released May 20, 2026, the OpenComputer framework replaces semantic guesswork with built-in monitoring:

1. Real-state verifier — connects directly to application backends; whether the agent actually wrote that line into a document is checked against the database, not inferred from pixels. 2. Self-evolving layer — the verifier itself learns from real execution feedback and grows stronger over time. 3. Realistic scenarios — replicates 33 desktop applications with 1,000 rigorous tasks.

The reward is computed from verified state, not opinion:

R_{reward} = \sum_{i} \omega_i \cdot \text{Verify}(s_{actual}, g_i)

Here the final reward depends on comparing the actual software state ( $s_{actual}$ ) against predefined goals ( $$g$$ ).

Reality check: heroes unmasked

Under OpenComputer's true-state verification, models that scored near-perfect on older benchmarks saw their scores cut roughly in half. Much of their earlier success was performance, not completion.

With verifiable environments like OpenComputer, AI assistants can finally graduate from typists to true task-completers.

---

Reference

Paper: OpenComputer: Verifiable Software Worlds for Computer-Use Agents
Released: May 20, 2026
arXiv: arXiv:2605.19769
Core contribution: solving hallucination and inflated scores in computer-use agent (CUA) evaluation caused by LLM-as-judge setups, via a 1,000-task auditable benchmark grounded in real application state.