06RESEARCH PAPERPreprint
Mercury Attention-Head Cartography and GQA/KV Sharing
發表日期
摘要
ABSTRACTZenodo 原文A Mercury methods paper treating an attention head as an addressable ownership contract: query slice, shared key/value owner, output slice, and route through nearby residual dimensions. The paper formalizes head cartography for GQA and MQA models where a visible head label is not enough. Copyright (c) 2026 Chen, Ho Yiing (norika). Licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).
全文 PDF
FULL TEXT1 個檔- 下載 PDFpaper.pdf・52 KB・本站
PDF 與 Zenodo 上的檔案相同(大小與 md5 已核對)。
論文資料
DETAILSZenodo 紀錄 20757833- 作者
- Chen, Ho Yiing
- 發表日期
- 資源類型
- Preprint
- DOI
- 10.5281/zenodo.20757833
- Zenodo
- zenodo.org/records/20757833
- 授權
- Creative Commons Attribution 4.0 International
作者 Chen, Ho Yiing 是織田紀香的本名(陳禾穎)英文拼寫。標題、摘要、關鍵字與授權照 Zenodo 原文。
關鍵字
KEYWORDS20 個- Mercury
- attention-head cartography
- GQA
- grouped-query attention
- KV sharing
- multi-query attention
- attention heads
- head ownership
- query key value
- transformer internals
- mechanistic interpretability
- LLM interpretability
- head contract
- KV cache
- attention visualization
- model cartography
- residual stream
- QKV mapping
- local LLMs
- neural atlas