研究人员开发了一个名为Anchored Learning的新框架，以减轻大型语言模型在监督微调过程中灾难性遗忘的问题。该方法通过使用动态移动锚点显式控制分布更新，该锚点在当前模型和冻结的参考模型之间进行插值。该方法在理论上保证了模型分布之间的稳定过渡，并在iGSM和MedCalc等基准测试中实证证明了性能下降的显著减少，同时保持了接近最优的收益。 AI
影响 解决了LLM中的灾难性遗忘问题，有望提高微调模型的稳定性和可靠性。
排序理由 该集群包含一篇arXiv预印本，详细介绍了一种稳定LLM微调的新方法。
AI 生成摘要 · Google Gemini · 来自 2 个来源。 我们如何撰写摘要 →
完整方法见我们的编辑标准。
arXiv:2605.04468v1 Announce Type: new Abstract: Post-training large language models (LLMs) often suffers from catastrophic forgetting, where improvements on a target objective degrade previously acquired capabilities. Recent evidence suggests that this phenomenon is primarily dri…
Post-training large language models (LLMs) often suffers from catastrophic forgetting, where improvements on a target objective degrade previously acquired capabilities. Recent evidence suggests that this phenomenon is primarily driven by excessive distributional drift during opt…