ALPHABET: Compact Sequence Model with Auditable Predictions

Sequence models are crucial for understanding temporal data, but often come with large parameter counts and black-box prediction mechanisms. Researchers have introduced ALPHABET, a novel, compact, and linear-time model designed to compress temporal history into stable complex pole modes. A key feature of ALPHABET is its explicitly auditable prediction interface, which provides transparency into its decision-making process. The model employs a direct bank to synthesize its modal states back into the feature trajectory, and an independent cascaded bank to analyze the transformed trajectory without resynthesis. An affine head then reads only modal energies and lag moments from both banks. This unique architecture allows ALPHABET to retain crucial temporal information, with each mode energy acting as a frequency-localized measurement of the second-order spectrum. Despite having only a few thousand parameters (6,437 at D=64 runtime anchor), ALPHABET achieves a mean rank of 3.97 across a comprehensive 82-task registry. It also delivers 5.02 times faster inference and 3.93 times faster complete training steps compared to nine baselines on average, demonstrating a powerful combination of efficiency, performance, and interpretability.