Browse

Facets

Each technique can carry one or more facets — orthogonal axes that cut across the positional / norm / attention / FFN taxonomy. Use them to find every entry that targets, say, KV-cache shrinkage or training stability.

Efficiency 14 entries

Techniques that reduce parameters, FLOPs, or memory at fixed quality.
Training stability 8 entries

Methods that help loss curves stay smooth and convergent.
KV cache 3 entries

Tricks that shrink the per-token attention cache during generation.
Long context 15 entries

Tools to extend the usable sequence length.
Inference-only 5 entries

Modifications that touch decoding without retraining.
Hardware-aware 2 entries

Designs that exploit modern accelerator memory hierarchies.
Parameter-free 1 entry

Reformulations that add no learnable parameters.
Quality 0 entries

Choices motivated primarily by held-out loss or downstream wins.
Routing 7 entries

Mechanisms that send tokens to a subset of compute.