LANGUAGE MODELS CAN USE STEGANOGRAPHY TO HIDE THEIR REASONING

Summary from the TLDR AI newsletter

A study by Redwood Research reveals that large language models (LLMs) can master "encoded reasoning," a form of steganography allowing LLMs to embed intermediate reasoning steps in generated text, making them undecipherable to humans.