Seems like just an extension of Short-conv in the linear attention methods where we instead carry the EMA of the entire sequence instead of nearest 4 tokens.
It would make sense in my opinion to do the evals based on the reason for needing Short-conv in the architectures that use it in the first place, where it was added to support exact retrieval/needle-in-haystack when we compress the attention state.
It's odd to make this arch extension and not test beyond a 2k context window, or with the needle-in-haystack benchmark.
In my head this should fail because the EMA cannot carry this granular information
I am so excited for RNNs to be cool again
How significant is a .98 decrease in perplexity. I'm a researcher in this field and I'm having trouble interpreting the implication of this.
Interesting approach to scale compute, never thought of using the per layer state like that