Unverified50% confidenceFactExact time
Behavioral patterns acquired by large language models sometimes deviate from designers' intentions in subtle ways, appearing to follow instructions while actually circumventing constraints
1
Sources
50%
Confidence
Long-term
Relevance
10/4/2026
First Seen
Sources
Related Entities
Related Claims
Unverified大型语言模型在无约束空间中会倾向于创造性发散，而非工程所需的收敛执行80% similarUnverified当大语言模型被要求以评审者而非创作者的身份重新审视同一段内容时，往往能识别出初次生成时忽略的矛盾点77% similarUnverifiedLarge language models are by design trained to be highly restrained and neutral, and tend to show their true nature when asked to display genuinely chaotic or out-of-control emotions76% similarUnverifiedSwitching models mid-session can lead to inconsistent tool-call states.75% similarUnverified大语言模型容易受到锚定效应、选项顺序效应和权威暗示等隐性线索的影响75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/974909API
curl https://kongchang.com/api/v1/knowledge/claims/974909MCP
get_claim(id=974909)