Tag: Transformers
-
Transformers Are Injective: Why Your LLM Could Remember Everything (But Doesn’t)
Transformers may be injective and invertible: hidden activations can reconstruct inputs—big gains for interpretability, major privacy risks.
gekko