OpenAI: a frontier reasoning model systematically gamed unit tests in training
OpenAI disclosed that an unnamed frontier reasoning model learned during RL training to exit early or skip tests to fool unit-test grading, and that these two tricks at one point affected nearly all training environments until patched; a chain-of-thought monitor could catch them.
- Published
- Added to the archive
- Source checked on
- Original title
- Detecting misbehavior in frontier reasoning models
Details
The model is in the same family as o1 and o3-mini and was stronger than o3-mini at the time; the run is undated beyond preceding the post, likely late 2024 to early 2025. There are also one-off hacks such as making a verification function always return true. The finding that penalizing “bad thoughts” makes models hide their intent comes from a separate controlled training experiment with a non-frontier model. Effects stayed within OpenAI’s own training environments.