Attention Heads Can't Compute Larger Parity Than They Have Count

Original: The Head Complexity of Boolean Functions in Single-Layer Attention

Source: arxiv.org ↗