06RESEARCH PAPERPreprint
Mercury Repository-to-Neuron Addressability
發表日期
摘要
ABSTRACTZenodo 原文A Mercury methods paper formalizing an address lift from code-side objects to model-side sites. The paper argues that an LLM atlas becomes more than a neuron picture when files, modules, layers, dimensions, heads, and measured sites remain linked in a testable chain. 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・53 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 20757832- 作者
- Chen, Ho Yiing
- 發表日期
- 資源類型
- Preprint
- DOI
- 10.5281/zenodo.20757832
- Zenodo
- zenodo.org/records/20757832
- 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
- repository-to-neuron addressability
- code-to-neuron mapping
- LLM interpretability
- code graph
- model internals
- neural atlas
- transformer internals
- mechanistic interpretability
- software provenance
- module addressability
- neuron address
- attention head address
- dimension-level address
- model cartography
- repo graph
- LLM tooling
- traceability
- interactive atlas
- local LLMs