When training LLMs to code with reinforcement learning, reward hacking is often detectable but difficult to prevent, since (a) environment hardening is incre…
机构：Google
来源：arXiv 2610.28707 | AI4Papers 论文推荐平台