Proprietary language models and their reasoning traces are vulnerable to API-based distillation, where an adversary trains a student model on the teacher mod…
机构:Meta
来源:arXiv 2610.37266 | AI4Papers 论文推荐平台