On-Policy Self-Distillation (OPSD) has emerged as a popular paradigm for large language model (LLM) self-improvement, allowing models to act as their own tea…
机构:UVA
来源:arXiv 2609.11699 | AI4Papers 论文推荐平台