Emergent AI Capabilities Linked to Learning Sparse Attention Patterns

New findings shed light on the mysterious emergence of capabilities in transformer language models, which often appear abruptly despite smooth improvements in pretraining loss. This research indicates that these emergent abilities, such as in-context learning, arise stochastically throughout the training process, with larger models tending to acquire them earlier. The key mechanism identified is the sudden learning of specific, task-relevant attention patterns. The study used synthetic datasets to demonstrate that the difficulty of learning these attention patterns is influenced by factors like context length and the sparsity of the pattern itself. It also found that increasing the number of attention heads improves learning efficiency, while increasing head dimension offers diminishing returns. This work provides a mechanistic explanation for emergence, linking it directly to the intrinsic challenge of transformers learning sparse attention patterns.