The model wasn’t update after the initial training, they’re definitely using in-context learning. The advancement here, as I understand it, is that they’ve achi

The model wasn’t update after the initial training, they’re definitely using in-context learning. The advancement here, as I understand it, is that they’ve achieved in-context learning with a very small model. From the paper: >> Few-shot instruction-learning task. MLC was evaluated on this task in several ways; in each case, MLC responded to this novel task through learned memory-based strategies, as its weights were frozen and not updated further.