It underlies many embedding and retrieval models, and CLM uses a bidirectional version to connect states with the actions actually taken. Its test value often falls as a predictable power law in compute, model size and data, which makes it useful for forecasting the value of scaling.