Sparse Attention

Routes inputs to subsets of parameters for scalable capacity.

Chooses which experts process each token.

Capabilities that appear only beyond certain model sizes.

Empirical laws linking model size, data, compute to performance.

Stores past attention states to speed up autoregressive decoding.

Encodes positional information via rotation in embedding space.

Encodes token position explicitly, often via sinusoids.

Techniques to handle longer documents without quadratic cost.