Proteus: Incremental Memory Activation for Long-Context Sequence Modeling
The quadratic cost of attention-based sequence models for long contexts has motivated a growing line of research on memory-based models that can compress context into a compact state. However, most existing memory models expose a static memory throughout the entire sequence. Because early tokens face no compression pressure, they occupy too many degrees of freedom and "pollute" the memory state,…
We haven't written up this one. arXiv cs.AI has the full story — the link below goes straight to it.
