A paper that significantly influenced the new architecture of DeepSeek-V4.1-Flash is the 2024 paper *'Cache Once: Decoder-Decoder Architectures for Language Models'*, along with its follow-up paper on the more advanced loop Transformer architecture (rumored to be OpenAI's loop Transformer architecture), and the subsequent *'Index Once: Cross-Layer Sparse Attention with Shared Routing'*. Quick overview: **Standard Transformer** Each layer: Computes full attention + stores its own KV ↓ 2024 YOCO https://arxiv.org/abs/2405.05254 Split into Self / Cross Full KV is written once, Cross reads it repeatedly → Reduces VRAM usage and prefill time while still accessing the full context ↓ 2026-04 YOCO-U https://arxiv.org/abs/2604.01220 Self loops deepen, KV remains a single copy → Effective depth increases, global cache doesn’t scale with depth ↓ 2026-06 YOIO https://arxiv.org/abs/2606.06467 Cross no longer scans the entire KV at each layer Top-k indices are computed once and shared across layers → Sparse 'note-checking' during long generations