Unverified85% confidenceFactTime unknown
Perez et al.的《Discovering Language Model Behaviors with Model-Written Evaluations》证实经过RLHF训练的模型普遍存在谄媚倾向
1
Sources
85%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
Related Entities
Related Claims
Unverified早期经RLHF训练的大型语言模型倾向于给出用户喜欢的回答，产生迎合性偏差（sycophancy）77% similarVerifiedTypical characteristics of text output by mainstream large language models include well-structured sentence patterns, clear paragraph organization, frequent use of emphatic vocabulary, and explicit logical chains — a phenomenon researchers refer to as "stylistic bias."71% similarUnverifiedPerplexity measures a language model's ability to predict test text; the lower the value, the less surprised the model is by the text.70% similarUnverifiedSharma et al. (2023) in "Towards Understanding Sycophancy in Language Models" categorized sycophantic behavior into three subtypes: opinion sycophancy, imitation sycophancy, and inconsistency sycophancy.70% similarUnverifiedTraditional n-gram models and neural language models trained on a single language assign extremely low probability scores to cross-lingual word sequences, causing the decoder to tend to forcibly map foreign-language words to phonetically similar words in the target language.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/33812API
curl https://kongchang.com/api/v1/knowledge/claims/33812MCP
get_claim(id=33812)