Phase-Associative Memory: A New Complex-Valued Sequence Model

Researchers have introduced Phase-Associative Memory (PAM), a complex-valued sequence model that operates in Hilbert space, departing from classical Boolean approaches. PAM accumulates outer products of complex token embeddings via conjugate inner products, aiming to capture the contextuality of semantic interpretation observed in humans and LLMs. The model was evaluated against a structurally matched real-valued ablation, with both architectures training stably. The work, detailed in arXiv:2604.05030v2, suggests that meaning is observer-dependent and indeterminate prior to interpretation, aligning with quantum logical mechanisms. This approach challenges traditional compositionality and separability assumptions in natural language processing.