The abstract opens by framing test-time memorization as a fundamental requirement for intelligent systems that must quickly store and retrieve sequential data. It points to the attention mechanism in large language models as a demonstration of this capability.
The proposed approach, named DeltaNet-PC, is presented as a way to achieve such memory without relying on backpropagation through time, a common training technique for recurrent architectures. The title suggests a focus on test-time behavior.
The provided abstract is truncated, so the specific architecture, experimental results, and comparisons are not available from this source. Readers interested in the full details would need to consult the complete paper.