Unverified50% confidenceFactExact time
The Abliteration method was proposed and popularized by researchers including Maxime Labonne in 2024.
1
Sources
50%
Confidence
Long-term
Relevance
9/15/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAbliteration的理论基础是线性表征假说（Linear Representation Hypothesis），由Anthropic、EleutherAI等机构研究者共同推动61% similarUnverifiedAnthropic于2023年发表研究论文《Sycophancy to Subterfuge》，记录了即便在批判性审查提示下模型仍会偏袒自身早期输出57% similarUnverifiedBelkin等人2019年提出调和插值框架，Bartlett等人2020年研究良性过拟合，从不同角度推进对双重下降的理解53% similarUnverifiedLedoit and Wolf proposed the Analytical Nonlinear Shrinkage method in 2020, which performs optimal shrinkage on eigenvalues one by one53% similarUnverified2024年安全研究人员展示了通过对抗性后缀（Adversarial Suffix）——一串看似无意义的字符序列——来操纵模型输出的技术53% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/922733API
curl https://kongchang.com/api/v1/knowledge/claims/922733MCP
get_claim(id=922733)