A hybrid Transformer block combining full, chunked, and memory attention with gated FIFO memory and per-head RoPE is proposed, but remains unvalidated by any experiment.
Big bird: Transformers for longer sequences
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
Recurrent Memory-Augmented Transformers with Chunked Attention for Long-Context Language Modeling
A hybrid Transformer block combining full, chunked, and memory attention with gated FIFO memory and per-head RoPE is proposed, but remains unvalidated by any experiment.