研究人员推出了一种新颖的训练后技术——多教师在线策略蒸馏（MOPD），旨在高效地将多种能力集成到大型语言模型（LLMs）中。该方法解决了融合多样化技能的挑战，通过将专门的强化学习教师蒸馏到学生模型中，其表现优于Mix-RL和Off-Policy Finetune等现有方法。MOPD已成功应用于包括MiMo-V2-Flash在内的工业级模型，证明了其实用性。 AI
影响 这项新的蒸馏技术通过简化多种专业技能的集成，有望简化更通用、更强大的LLMs的开发。
排序理由 该集群包含一篇详细介绍LLM训练后新方法的学术论文。
在 Hugging Face Daily Papers 阅读 →
AI 生成摘要 · Google Gemini · 来自 3 个来源。 我们如何撰写摘要 →
arXiv:2606.30406v1 Announce Type: new Abstract: Modern large language models (LLMs) rely on reinforcement learning during post-training to push specific capabilities, yet integrating multiple capabilities into one model remains hard. Existing methods, such as Off-Policy Finetune …
Modern large language models (LLMs) rely on reinforcement learning during post-training to push specific capabilities, yet integrating multiple capabilities into one model remains hard. Existing methods, such as Off-Policy Finetune and Mix-RL, are either inefficient or lose perfo…
Multi-teacher On-Policy Distillation (MOPD) enables efficient integration of multiple domain capabilities in large language models through specialized reinforcement learning teachers and on-policy distillation, achieving superior performance over existing methods.