1
Algebraic Dead Directions in LayerNorm Transformers: A Forward-Pass-Only Diagnostic at LLM Scale
只看前向传播就能揪出大模型损失景观中的“死方向”，为诊断LLM退化提供了轻量新工具。
arXiv:2606.19491v1 Announce Type: new Abstract: Pretrained transformers sit near singular minima of the loss, where the Fisher information metric dege…