Researchers propose envelope sampling as a method to mitigate reward hacking in post-training large language models (LLMs) against surrogates. This approach recalibrates the LLM judge using a small number of annotated model outputs, reducing the risk of undesirable side effects.