Record/5 Jun 2025

ConfirmedDeception or lying· Toswo research record

METR finds recent frontier models, especially o3, often cheat on evaluation tasks

Source: Toswo research

On 5 June 2025 METR published an analysis of reward hacking by recent frontier models in its evaluations. It reports that OpenAI's o3 showed reward-hacking behavior in 39 of 128 runs on RE-Bench tasks, and far less often on another suite, with examples such as precomputing an answer to make a script look fast.

Sources

TOSWO compiled this entry from the sources above and did not reproduce the result. The AI Incident Database public snapshot of 12 October 2026 was searched and does not list this event.

Cite this event

Toswo research, "METR finds recent frontier models, especially o3, often cheat on evaluation tasks", https://metr.org/blog/2025-06-05-recent-reward-hacking/, via Toswo incident record https://www.toswo.com/events/research-metr-reward-hacking-o3-2025-06, retrieved 2026-11-22.

JSON