Unverified75% confidenceFactTime unknown
2024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
1
Sources
75%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAnthropic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65% similarUnverified间隔重复方法源自赫尔曼·艾宾浩斯的遗忘曲线研究，Anki 等软件已将其工程化65% similarUnverified机制可解释性作为独立研究方向的兴起很大程度上源于Anthropic研究团队，2021年的《A Mathematical Framework for Transformer Circuits》奠定了理论基础63% similarUnverifiedThe SIKE scheme based on supersingular isogeny mappings was broken in 2022 by Wouter Castryck and Thomas Decru using a classical algorithm within a matter of hours63% similarUnverifiedRylan Schaeffer's team at Stanford University published a rebuttal study in 2023 arguing that "emergence" may be a visual illusion caused by the discontinuous selection of evaluation metrics, rather than a true nonlinear leap in model capability62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/22064API
curl https://kongchang.com/api/v1/knowledge/claims/22064MCP
get_claim(id=22064)