A measure-zero reminder: Transformers can be almost surely injective

CatAstro_Piyush · x · 2026-07-27

This post argues that measure zero is often misunderstood in ML and notes that there can still be countably infinite counterexamples while a theorem remains true almost surely.

It ties that intuition to a result about Transformers: for finite-length prompts, the set is countable, and the cited work claims Transformers are almost surely injective, meaning the exact input text can in principle be recovered from hidden states under almost all parameter settings.

Original post →

More from Research

Research channel →