Giorgos Nikolaou, Tommaso Mencattini, Donato Crisostomi, Andrea Santilli, Yannis Panagakis, and Emanuele Rodolà, from EPFL, the University of Athens, the University of Rome, and the Archimede Research Center, publish the scientific paper: “Language Models are Injective and Hence Invertible.” The “prompt => LLM answer” function is invertible. AI is moving from a practice to a science, with the necessary mathematical framework. Contrary to popular belief that LLMs destroy information as data passes through their layers, this apparently isn’t the case! This means we can reconstruct the EXACT input provided to an LLM simply by observing its internal state. This has implications for privacy, security, and our understanding of model behavior.
