Looks very exciting! It always seemed likely that the optimal model architecture would be one forward pass per logical step. LLMs are quite wasteful on one end of the spectrum at one forward pass per token. This looks more like one forward pass per prompt? Would be very interesting if one could somehow chain this together now into multi-step (i.e. multi forward pass) reasoning, and optionally in the end unpack cheaply with small LLM.