Decoder doesn't cross-attend to encoder KVs — just periodically reads final encoder states
stochasticchasm · x · 2026-09-11
Reading the architecture blog closely, the author notes the CED decoder does not attend to encoder-section KVs across multiple layers — it just periodically reads the final encoder states, which makes sense as those are the richest and follows the original Transformer. It's also not true cross-attention: the self-attention simply has prefilled KVs. The author adds that the team already has a strong stable-RL recipe (batch-invariant kernels, low-precision inference), so data deserves attention now, and predicts value modeling will eventually appear.
Related event: Leaked DeepSeek V4.1 Flash Tech Report Reveals KV Cache-Obsessed Flagship(44 posts)→
More from Models
- MissionReport benchmark: only 3 of 540 AI-generated government reports pass all criteria — Exp_Mark · 2026-10-27
- Local beats cloud: Qwen 3.8 FN handles private student data as well as frontier models — dr_lm · 2026-10-27
- Aidan Gomez: model's CoT said no tool calls right before making them — aidangomez · 2026-10-27
- Nat Lambert: Cohere's model is narrow RL training on internal infra ill-suited for open weights — natolambert · 2026-10-27
- OpenAI teases DOTS is coming to Europe 'soon' — mark_k · 2026-10-27
- Clifford Pickover challenges AI to decipher a message he encoded in Norse runes — pickover · 2026-10-27