We propose Memory-Integrated Reconfigurable Adapters, a unified ML framework integrating Hopfield-style associative memories atop a shared backbone, demonstrating remarkable flexibility across domain shifts and sequential task exposures.

Memorization in transformer LMs is tied to pattern acquisition. It is non-trivial and happens in bursts according to shared patterns. Intriguingly, the relative memorization speed of larger and smaller models can change based on the pattern type.

Steering given distributions towards ideal distributions, where fairness and accuracy are not at a tradeoff.