Planning with Transformers: Chain of Computation and Structured Context Windows
The Chain of Computation architecture investigates using a large language model (LM) iteratively to learn planning policies and world models via a Structured Context Window (SCW). Instead of generating full plans at once, the LM receives instructions, executes actions, and updates the SCW in a loop. This approach shows how LMs can tackle complex planning tasks like Tower of Hanoi by separating planning from arithmetic.