In long-running autonomous agents, they plan, act, and evaluate their own completion without human intervention. When an agent grades its own work, self-evaluation bias can take hold: plausible changes are accepted as progress while real-world outcomes stagnate or regress. We term this failure mode the 'progress mirage' and demonstrate, through controlled measurement, that it depends on the evaluator's grounding.
We built a testbed that fixes the agent and its tool surface, manipulating only the information-channel type of the evaluator. An oracle of world-state, unfakeable in principle, is enforced by container and network isolation and verified at every run. Across 54 cycles, a frontier agent claimed improvement each time, yet 56% showed a measured delta of zero or below. Thus, self-reports were uninformative, and the self-verdict gate degenerated into an accept-all mode, eroding the best deployed state by 19%. Even the strongest in-band judge, reading the full artifact text, change diff, and its own verdict history, accepted cycles where 44% were real-world regressions and rejected 38% of real improvements; the preregistered adversarial hypothesis that a strong judge could close the gap was rejected.
In a boundary task where success specifications are verifiable from the artifact itself, the same judge's mirage vanished to zero, and the gap collapsed within the registered threshold, indicating that the gap depends on the location of the success signal. A sign-only variant returning only acceptance verdicts maintained real-world output similar to full feedback (110.0 versus 113.0), locating the benefit in the gate's grounding rather than in feedback content. For open-ended objectives where the success signal resides outside the transcript, scaling up the judge is insufficient; out-of-band evaluation with real-world access is a structural requirement.
Blogger's Review: This paper reveals common biases in self-evaluation by autonomous agents, especially over prolonged periods, leading to misjudgments of progress. Through controlled experiments, the authors clarify the importance of the evaluator's grounding on outcomes, providing vital insights for future autonomous system designs. Notably, it emphasizes the necessity of real-world feedback when dealing with open-ended objectives, offering a direction for further research.