Chain-of-thought can also be interpreted as a simple external memory tool: The model writes to a scratchpad, then reads it, and answers the question. The scratchpads to which the model outputs text and inputs text are powerful tools that need further exploration.

→ View original post on X — @nandodf,