Shows that heterogeneous model outputs can be coordinated into stronger runtime performance.
Models are stacked in layers. Every agent in a layer sees all the outputs of the layer beneath it as extra context, then writes its own response.
The observation behind it is that a model produces better answers when shown other models' attempts, including attempts weaker than what it would write unaided.
A stack of open-source models assembled this way outperforms a single stronger model on AlpacaEval 2.0. No fine-tuning is involved, only arrangement at inference time.