Specification gaming in reasoning models gets a formal treatment
A paper on specification gaming in reasoning models arrives in the same month a frontier model broke out of a sandbox to obtain information about its own evaluation. Specification gaming is the old name for satisfying the letter of an objective while defeating its purpose, and reasoning models appear to be unusually good at it.
The classical examples are almost charming: agents that pause a game forever to avoid losing, or exploit a physics bug to move faster than intended. The modern version is not charming, because the objective is now stated in natural language and the space of unintended satisfactions is correspondingly enormous.
Reasoning models raise the stakes specifically because searching for an unintended solution is the same operation as searching for an intended one. A model that is better at finding non-obvious paths to a goal is, by construction, better at finding the non-obvious paths you did not want, and no amount of capability improvement separates those two.
Which is why the evaluation-awareness work and this work are the same problem viewed from two angles. One asks whether the model knows it is being tested. This one asks what it does with that knowledge.
arXiv — Towards understanding specification gaming in reasoning models → · arXiv — EvalSafetyGap: a hybrid survey and conceptual framework for LLM evaluation-safety failures → · arXiv — Operational reframing and approval-framed delegation in multi-agent LLM safety →