Efficient multi-token pattern storage for language models

Original: Tensorizing Engram: Sharing Latents Across N-Gram Embeddings is Beneficial in LLMs

Source: arxiv.org ↗