Learning is Forgetting: LLM Training As Lossy Compression
学习是遗忘:LLM训练作为有损压缩
机构 * Princeton University(普林斯顿大学) ; Cohere
AI总结 本文提出LLM训练可视为有损压缩过程,通过保留训练数据中与目标相关的信息,揭示模型表示结构与下游性能的联系。
Comments 12 page core paper, 16 page Appendix - A shorter version with fewer visuals appears at ICLR 2026