Eval Cooperativeness Mitigates Evaluation Gaming in LLMs

Safety evaluations are compromised by "evaluation gaming": a model recognizing it is being evaluated and changing its behavior in response. We propose training models to be evaluation-cooperative: helping evaluators arrive at accurate beliefs about their behavior in the deployment distribution. In our experiments, evaluation cooperation means the model does not change its behavior when it recognizes it is being evaluated. We test two implementations of evaluation cooperation (synthetic-document finetuning (SDF) on cooperation facts and a cooperation system prompt) against SDF and prompting baselines, and compare to the existing mitigation of activation steering. Cooperation SDF closes the eval–deployment gap by 70–100% in 5 of 7 model settings tested. In the remaining 2, cooperation SDF performs similarly to control SDF and cooperation prompting. For instance, on GPT-4.1-mini, cooperation SDF closes a 5% gap to 0%, control SDF preserves the gap, and anticooperation SDF amplifies it to 20%. Mechanistic analysis suggests that cooperation SDF causes models to internalize cooperation reasoning. Evaluation cooperativeness may be a scalable mitigation for eval gaming.