Artem Karpov

New research explores LLMs' ability to hide reasoning within text

Researchers have explored the concept of steganographic chain-of-thought (CoT) in large language models, where models hide intermediate reasoning within innocuous text to evade monitoring. Experiments across 34 models, …