Unverified50% confidenceFactExact time
大语言模型存在奉承偏差（Sycophancy Bias），天然倾向给出正面高分评价，这一偏差在小模型中尤为突出
1
Sources
50%
Confidence
Long-term
Relevance
7/12/2026
First Seen
Sources
Local AI Agent Reads arXiv So You Don't Have To: Only the Papers That Actually Matter
redditr/ollama7/10/2026
Related Claims
UnverifiedSycophancy bias refers to the tendency of models to give answers that the user seems to want to hear.71% similarUnverifiedResearch has found that user preferences may be biased toward longer, more elaborately formatted responses, a phenomenon known as verbose bias or length bias.69% similarUnverifiedLLM-as-Judge存在位置偏见（Position Bias）和奉承偏见（Sycophancy Bias）等问题69% similarUnverifiedSingle judge models have known biases including position bias, length bias, and self-enhancement bias.69% similarUnverifiedThe Courtesy Bias refers to the tendency of test participants to give more positive feedback when they have a social relationship with the researcher.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/491243API
curl https://kongchang.com/api/v1/knowledge/claims/491243MCP
get_claim(id=491243)