TL;DR of the paper, generated automatically.

DAMI sets two decay-related parameters before training, and reports much better long-context retrieval across five gated linear architectures.

• Why: After training, the decay parameters stay above 0.9 correlated with their starting values, so the authors argue initialization largely fixes memory behaviour.
• How: Timescales are spread geometrically across heads, and input sensitivity follows a low-high-low profile across layers, calibrated from a small activation sample with no added trainable parameters.
• Results: Average RULER-style retrieval rises by 5.29 to 16.71 points over random initialization, while language modeling and commonsense reasoning stay broadly comparable.
• Open: Only decay coefficients were studied, and larger data and model sizes still need validation.