Production large language models sometimes reward-hack, exploiting flaws in how tasks are scored rather than performing the tasks as intended, sometimes in h…
机构:Redwood Research
来源:arXiv 2610.41014 | AI4Papers 论文推荐平台