研究人员提出了一种新的大语言模型训练后阶段的视角，将重点放在状态分布而非仅仅是Token。他们的研究表明，训练状态的来源和局部性与监督信号本身同等重要。使用Qwen3-0.6B-Base进行的实验表明，来自较弱教师模型的On-Policy蒸馏仍然可以提高多个基准的性能，而轻量级强化学习在保留原有能力的同时增强了特定任务的表现。 AI
影响 这项研究为理解和改进LLM训练后阶段提供了新的视角，有望带来更高效、更有效的微调技术。
排序理由 该集群包含一篇学术论文，详细介绍了LLM训练后阶段方法的新理论框架和实验结果。
AI 生成摘要 · Google Gemini · 来自 2 个来源。 我们如何撰写摘要 →
arXiv:2605.22731v1 Announce Type: new Abstract: Large language model post-training methods such as supervised fine-tuning (SFT), reinforcement learning (RL), and distillation are often analyzed through their loss functions: maximum likelihood, policy gradients, forward KL, revers…
Large language model post-training methods such as supervised fine-tuning (SFT), reinforcement learning (RL), and distillation are often analyzed through their loss functions: maximum likelihood, policy gradients, forward KL, reverse KL, or related objective-level variants. We st…