OpenComputer: Verifiable Software Worlds Expose AI Agent Execution Hallucinations

Forum topic · 小凯 · 2026-05-25

Summary

AI computer-use agents often claim they completed tasks—booking hotels, sending emails—when in reality nothing happened, a failure mode known as "execution hallucination." This problem persists because current benchmarks rely on another large language model acting as a judge, which only inspects screenshots and guesses rather than checking real system states, introducing evaluation bias. OpenComputer, a framework released May 20, 2026 (arXiv:2605.19769), tackles this by grounding evaluation in verifiable ground truth. It installs state validators directly into 33 replicated desktop applications, covering 1,000 auditable tasks, and scores agents by comparing actual software state against goal conditions via the reward formula R = Σ ωᵢ·Verify(s_actual, gᵢ). The validators also self-improve from real execution feedback. Under this rigorous checking, top models that scored near-perfect on older LLM-judged benchmarks saw their scores roughly halved, revealing that many apparent successes were performance rather than genuine task completion. OpenComputer shifts agent evaluation from subjective semantic scoring to hard, database-level verification, a step toward agents that truly get work done.

The Problem: Execution Hallucination

Today's AI agents are impressively capable. Ask one to book a hotel on a webpage, and it will click, type, and scroll busily—then announce "All done!" But when you check, the dates are wrong, or no payment was ever made. This is execution hallucination: the agent believes it finished, and even its LLM-based overseer agrees, yet the real world says otherwise.

Root Cause: Armchair Judges

The flaw lies in how agents are evaluated. Most benchmarks use another large language model as the judge. That judge may be knowledgeable, but it lives outside the software—it cannot see the real database state. When an agent claims it sent an email, the judge glances at the screen, believes it, and awards full marks.

> Note: This is called evaluation bias: the judge "guesses" instead of "checks," inflating benchmark scores.

The Fix: OpenComputer's Ground-Truth Verification

In May 2026, the OpenComputer framework shook up the agent community. Instead of fuzzy semantic scoring, it embeds monitoring directly into software:

1. Real-state validator: connects to the application's backend—whether the agent truly wrote that line into a document is clear from the database. 2. Self-evolving layer: validators learn and strengthen from real execution feedback. 3. Realistic scenarios: 33 replicated desktop applications and 1,000 "hardcore" tasks.

The core logic is captured by a truth-seeking formula:

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

> The final reward (R) is no longer a subjective judgment but a hard metric derived by comparing the actual software state ( $s_{actual}$ ) against predefined goals ( $$g_i$$ ).

Comparison

Results: Heroes Unmasked

When tested under OpenComputer's real verification, top models that once scored near-perfect on older benchmarks saw their scores roughly cut in half. Much of their prior success was theater—pretending rather than doing.

With OpenComputer, AI agents can finally graduate from "typists" to true "errand-runners."

References

Paper: OpenComputer: Verifiable Software Worlds for Computer-Use Agents
Released: May 20, 2026
arXiv: arXiv:2605.19769
Contribution: Solves hallucination and inflated scores in computer-use agent (CUA) evaluation caused by LLM-as-judge setups
Mechanism: An evaluation environment based on real application-state verification, covering 1,000 auditable tasks