NORIKA ODA

06RESEARCH PAPERPreprint

Coverage Audits for LLM Interpretability Claims

Chen, Ho Yiing

發表日期

DOI 10.5281/zenodo.20729124　Zenodo 紀錄頁　PDF（48 KB）

01

摘要

ABSTRACTZenodo 原文
A Mercury methods note arguing that sparse probing, activation patching, subspace intervention, and pruning claims should report a coverage audit before making broad localization claims. The audit uses 19 fullgrid manifests and shows a 1/7 pass rate for older sparse or coarse protocols versus 12/12 for per-channel quantile dense runs.
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 20729124
作者
Chen, Ho Yiing
發表日期
資源類型
Preprint
DOI
10.5281/zenodo.20729124
Zenodo
zenodo.org/records/20729124
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
  • coverage audits
  • LLM interpretability
  • mechanistic interpretability
  • activation coverage
  • sparse probing
  • fullgrid
  • coverage failure
  • activation patching
  • model pruning
  • measurement validity
  • interpretability claims
  • cell coverage
  • dense activation measurement
  • sampling bias
  • neural atlas
  • model evaluation
  • result governance
  • transformer internals
  • reproducibility
05

其他研究論文

MORE