Chorus: Efficient Layer-wise Adaptation with Shared N-gram Embeddings
November 16, 2026
Chorus uses a single set of shared n-gram embedding tables for an entire network, with each layer applying its own projection to capture layer-specific preferences. This approach increases model capacity with minimal memory overhead compared to separate per-layer tables.
HOW THIS AFFECTS YOU
●
builderThis provides a way to improve model performance with low computational and memory impact.
●
researcherYou can add n-gram capacity to models without the massive memory cost of per-layer tables.