MultiHashFormer Enables Hash-Based Autoregression to Cut Embedding Parameters

MultiHashFormer solves the many-to-one collision problem that blocked hash-based embeddings from causal language models. Each token gets a unique hash signature from multiple independent hash functions, compressed into a latent vector by a Hash Encoder, with a Hash Decoder reconstructing the next token's signature. This reduces embedding matrix size without sacrificing token uniqueness in generative settings.