LET'S THINK DOT BY DOT

Summary from the TLDR AI newsletter

This paper shows that chain-of-thought (CoT) thinking can be obfuscated with ‘...' tokens. This requires training the model but shows that interpreting the CoT steps may not be possible and thinking can be hidden by the model.