NORIKA ODA

06RESEARCH PAPERPreprint

Mercury Model Birthmarks: Structural Signatures of Token Cost, Latency, and Hidden Defects Across LLM Scales

Chen, Ho Yiing

發表日期

DOI 10.5281/zenodo.20757828　Zenodo 紀錄頁　PDF（53 KB）

01

摘要

ABSTRACTZenodo 原文
A Mercury methods paper defining model birthmarks as stable structural signatures of tokenizer burden, latency habits, routing scars, dead-tissue regions, alignment reflexes, and compression scars. The paper argues that parameter count is not a health report and proposes measurable birthmark vectors for language model inspection. Copyright (c) 2026 Chen, Ho Yiing (norika). Licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).
02

全文 PDF

FULL TEXT1 個檔

PDF 與 Zenodo 上的檔案相同（大小與 md5 已核對）。

03

論文資料

DETAILSZenodo 紀錄 20757828
作者
Chen, Ho Yiing
發表日期
資源類型
Preprint
DOI
10.5281/zenodo.20757828
Zenodo
zenodo.org/records/20757828
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。

04

關鍵字

KEYWORDS20 個
  • Mercury
  • model birthmarks
  • LLM diagnostics
  • structural signatures
  • language model interpretability
  • token birthmarks
  • latency birthmarks
  • routing scars
  • dead tissue
  • alignment reflex
  • compression scars
  • model health
  • transformer internals
  • mechanistic interpretability
  • model compression
  • tokenization defects
  • LLM evaluation
  • local LLMs
  • model anatomy
  • hidden defects
05

其他研究論文

MORE